AIMFP

AIMFP: AI Modular Functional Procedural Programming

[PyPI]

A language-agnostic programming paradigm designed for AI-generated and AI-maintained codebases

Table of Contents

• What is AIMFP?
• Core Principles
• Architecture Overview
• Database Architecture
• How It Works
• Getting Started
• Platform Support
• Directives System
• Project Lifecycle
• Example Workflow
• Documentation
• Development
• Design Philosophy
• Features
• Usage Examples
• Privacy Policy
• Support
• License

What is AIMFP?

AIMFP (AI Modular Functional Procedural) is a programming paradigm that combines:

• Pure functional programming principles (referential transparency, immutability, composability)
• Procedural execution patterns (explicit sequencing, no hidden state)
• Database-driven project management (persistent state, instant context retrieval)
• Directive-based AI guidance (deterministic workflows, automated compliance)

Two Ways to Use AIMFP

Use Case 1: Regular Software Development

• Build applications (web apps, libraries, CLI tools, etc.)
• AIMFP enforces FP compliance and manages your project
• You write code, AI assists with FP standards and project tracking
• Example: Building a web server, calculator library, data processor

Use Case 2: Custom Directive Automation

• Define automation rules (home automation, cloud management, workflows)
• AIMFP generates and manages the automation codebase for you
• You write directive definitions (YAML/JSON/TXT), AI generates the implementation
• The project's code IS the automation code generated from your directives
• Example: Smart home control system, AWS infrastructure manager, workflow automator

Key Principle: One AIMFP instance per project directory. You would NOT mix a web app with home automation directives. Run separate instances for separate purposes.

Why AIMFP?

Traditional programming paradigms were designed for humans. AIMFP is optimized for AI-human collaboration:

Core Principles

1. Functional-Procedural Hybrid

# ✅ AIMFP-Compliantdef calculate_total(items: List[Item]) -> float:    """Pure function: deterministic, no side effects"""    return reduce(lambda acc, item: acc + item.price, items, 0.0)
# ❌ Not AIMFP-Compliantclass Calculator:    def __init__(self):        self.total = 0  # Hidden state
    def add_item(self, item):        self.total += item.price  # Mutation

2. Database-Indexed Logic

Every function, file, and dependency is tracked in SQLite. AI accesses this data through helper tools — not raw SQL:

AI calls: get_functions_by_file(file_id)→ Returns: all functions in that file with name, purpose, parameters, returns

No source code reparsing required. Instant context retrieval across sessions.

3. AI-Readable Code

• Flat structure: No deep inheritance hierarchies
• Explicit dependencies: All parameters passed explicitly
• Pure functions: Same inputs → same outputs
• Metadata annotations: Machine-readable function headers

4. Finite Completion Paths

Project: MatrixCalculator├── Completion Path (3 stages)│   ├── 1. Setup (completed)│   ├── 2. Core Development (in progress)│   │   ├── Milestone: Matrix Operations│   │   │   ├── Task: Implement multiply│   │   │   ├── Task: Implement transpose│   │   │   └── Task: Add validation│   │   └── Milestone: Vector Operations│   └── 3. Finalization (pending)

5. Language-Agnostic

AIMFP works with Python, JavaScript, TypeScript, Rust, Go, and more. FP directives adapt to language-specific syntax while maintaining universal standards.

Architecture Overview

┌─────────────────────────────────────────────────────┐│            AI Assistant (Claude, GPT-4, etc.)        ││  - Receives natural language commands                ││  - Follows directives (FP baseline + project mgmt)   ││  - Calls MCP tools to read/write databases           ││  - Generates FP-compliant code                       │└────────────────────┬────────────────────────────────┘                     │ MCP Protocol (JSON-RPC over stdio)┌────────────────────▼────────────────────────────────┐│                 MCP Server                           ││  - Exposes helper tools via JSON-RPC                  ││  - Manages four-database connections                 ││  - Provides CRUD helpers for all databases           ││  - No business logic — AI makes all decisions        │└───┬────────────────────┬─────────────────────────┬──┘    │                    │                         │┌───▼──────────────┐ ┌───▼────────────────┐ ┌─────▼─────────────────┐ ┌──────▼────────────────┐│  aimfp_core.db    │ │  project.db        │ │  user_preferences.db  │ │  user_directives.db   ││  (Global,        │ │  (Per-Project,     │ │  (Per-Project,        │ │  (Per-Project,        ││   Read-Only)     │ │   Mutable)         │ │   Mutable)            │ │   Optional)           ││                  │ │                    │ │                       │ │                       ││ - FP directives  │ │ - Project metadata │ │ - Directive prefs     │ │ - User directives     ││ - Project mgmt   │ │ - Files & funcs    │ │ - User settings       │ │ - Execution stats     ││ - User prefs     │ │ - Task hierarchy   │ │ - AI learning log     │ │ - Dependencies        ││ - User systems   │ │ - Themes & flows   │ │ - Tracking features   │ │ - Generated code refs ││ - Helper defs    │ │ - Completion path  │ │ (All opt-in)          │ │ - Source file tracking││ - Directive flow │ │ - Runtime notes    │ │                       │ │ (Logs in files)       │└──────────────────┘ └────────────────────┘ └───────────────────────┘ └───────────────────────┘

Use Case 2 adds a second path that never touches the MCP server. Your automation has to run when no AI session is open, so it imports AIMFP as a plain library instead:

┌──────────────────────────────┐        ┌──────────────────────────────┐│  cron / systemd / your daemon │        │  AI Assistant (a session)     ││  runs forever; AIMFP does not │        │  present only when you are    │└───────────────┬──────────────┘        └───────────────┬──────────────┘                │ starts                                 │ MCP┌───────────────▼──────────────┐        ┌───────────────▼──────────────┐│  runner.py  (YOUR project)    │        │         MCP Server            ││  generated by AIMFP, owned    │        │                               ││  and editable by you          │        │  check_directive_health       ││                               │        │  get_directive_execution_stats││  import aimfp.hooks  ◄────────┼── a    │  get_recent_directive_errors  ││  get_due_directives(...)      │  library                              ││  run_directive(...)           │  call, └───────────────┬──────────────┘└───────────────┬──────────────┘  not RPC               │ reads                │ writes                                 │                └────────────►  user_directives.db  ◄────┘                                + .aimfp-project/logs/

The runner writes; the AI reads it back later. AIMFP itself is inert in between — it has no timer and starts nothing.

Database Architecture

aimfp_core.db (Global, Read-Only)

Location: Within MCP server installation directory (user-defined location, configured in AI client)

Purpose: Immutable knowledge base containing all AIMFP standards, directives, and helper definitions.

Key Tables:

• directives: All FP, project, and user preference directives (workflows, keywords, thresholds)
• helper_functions: Database, file, Git, and FP utilities organized across multiple registry files
• directive_helpers: Many-to-many junction table mapping directives to their helper functions with execution metadata
• categories: Directive groupings (purity, immutability, task management, etc.)
• directive_flow: Status-driven directive navigation and routing

Helper-Directive Relationship (New in v1.4):

• One directive can use many helpers, one helper can serve many directives
• Junction table stores: execution context, sequence order, parameter mappings
• Enables flexible helper reuse and clear execution flow
• Defined in directive_helpers junction table in aimfp_core.db

Helper Classification:

• Tool: All helpers are exposed as MCP tools (AI calls directly via MCP)
• Sub-helper (is_sub_helper = TRUE): Internal utility called by other helpers only (not exposed to AI)

Helper Registry (Development Staging):

• Helper definitions maintained in dev/helpers-json/*.json during development
• Developers modify JSON files, then import to database when complete
• Each helper includes used_by_directives field for relationship mapping
• Database import script populates aimfp_core.db from JSON files before release
• Production: Users query aimfp_core.db (pre-populated), NOT JSON files
• JSON files are dev-only staging area, never shipped with package

Read-Only Philosophy: This database is version-controlled and immutable once deployed. AI reads from it but never modifies it.

project.db (Per-Project, Mutable)

Location: <project-root>/.aimfp-project/project.db

Purpose: Persistent state for a single AIMFP project. Tracks code structure, tasks, and runtime notes.

Key Tables:

• project: High-level metadata (name, purpose, goals, status, user_directives_status, last_known_git_hash)
• files, functions, interactions: Code structure tracking
• themes, flows: Organizational groupings
• completion_path, milestones, tasks, subtasks, sidequests, items: Hierarchical roadmap
• task_files: Files worked on for each task/subtask/sidequest — linked automatically while the item is in progress
• modules, module_files, file_flows: Reusable code boundaries and file-to-flow assignments
• notes: Runtime logging with optional directive context (source, severity, directive_name)
• types: Algebraic data types (ADTs)
• infrastructure: Project setup (language, packages, testing)
• work_branches: Git collaboration metadata (user, purpose, merge strategy)
• merge_history: FP-powered merge conflict resolution audit trail

User Directives Integration: The project.user_directives_status field tracks whether user directives are initialized (NULL/in_progress/active/disabled), allowing aimfp_run and aimfp_status directives to include user directive context when active.

Enhanced Notes: The notes table now includes source (user/ai/directive), directive_name (optional context), and severity (info/warning/error) for better traceability.

user_preferences.db (Per-Project, Mutable)

Location: <project-root>/.aimfp-project/user_preferences.db

Purpose: User-specific AI behavior customizations and opt-in tracking features.

Key Tables:

• directive_preferences: Per-directive behavior overrides (atomic key-value structure)
• user_settings: Project-wide AI behavior settings
• tracking_settings: Feature flags for opt-in tracking (all disabled by default)
• ai_interaction_log: User corrections and learning data (opt-in)
• fp_flow_tracking: FP compliance history (opt-in)
• issue_reports: Contextual bug reports (opt-in)

Cost Management Philosophy: All tracking features disabled by default to minimize API token usage. Project work should be cost-efficient; debugging and analytics are opt-in.

User Customization Example:

-- User says: "Always add docstrings"INSERT INTO directive_preferences (directive_name, preference_key, preference_value)VALUES ('project_file_write', 'always_add_docstrings', 'true');
-- Next file write automatically includes docstrings

user_directives.db (Per-Project, Optional)

Location: <project-root>/.aimfp-project/user_directives.db

Purpose: Store user-defined domain-specific directives for automation (home automation, cloud infrastructure, etc.). When this database exists, the AIMFP project IS dedicated to building and managing the automation code generated from these directives.

This database only exists in Use Case 2: Custom Directive Automation projects. In regular software development projects, this database is not created.

Key Tables:

• user_directives: Directive definitions (triggers, actions, status, validated configuration). trigger_config and action_config have an enforced grammar — see Directive Configuration Schema
• directive_executions: Execution statistics (summary only, detailed logs in files)
• directive_dependencies: Required packages, APIs, environment variables
• directive_implementations: Links directives to generated code files
• helper_functions: AI-generated helper functions (project-specific implementation utilities)
• directive_helpers: Many-to-many junction table mapping user directives to their helpers
• source_files: Tracks user directive source files (YAML/JSON/TXT)
• logging_config: File-based logging configuration

Helper Functions (New in v1.0):

• AI generates project-specific helper functions for user directives
• Tracked with implementation status: not_implemented → generated → tested → approved
• Enforces FP compliance (pure functions) for all generated code
• Same many-to-many relationship pattern as aimfp_core.db

File-Based Logging Philosophy: Database stores state and statistics only. Detailed execution logs (30-day retention) and error logs (90-day retention) are stored in rotating files at .aimfp-project/logs/.

Directory Structure Comparison:

# Use Case 1: Regular Software Developmentmy-web-app/├── src/                    # Your application code├── tests/                  # Your tests└── .aimfp-project/          # AIMFP tracks your application    ├── project.db    └── user_preferences.db
# Use Case 2: Custom Directive Automationhome-automation/├── directives/             # ← User writes directive files here│   ├── lights.yaml│   └── security.yaml├── src/                    # ← AIMFP GENERATES this code│   ├── lights_controller.py│   ├── security_monitor.py│   └── runner.py           # ← the RUNNER: your scheduler calls this,│                           #    it imports aimfp.hooks as a library├── tests/                  # ← AIMFP GENERATES tests└── .aimfp-project/          # ← AI-managed only, user never touches    ├── project.db          # Tracks generated src/ code    ├── user_preferences.db    ├── user_directives.db  # References ../directives/ files    └── logs/               # 30/90-day execution logs
# ...and outside the project entirely:/etc/cron.d/home-automation  # ← what actually STARTS runner.py. Not AIMFP.

Example Workflow (Automation Project):

1. User creates directives/lights.yaml in their project
2. User tells AI: "Parse my directive file at directives/lights.yaml"
3. AI parses and validates through interactive Q&A
4. AI generates FP-compliant implementation code in src/
5. AI tracks generated code in project.db (files, functions, tasks)
6. AI generates a runner and wires up your scheduler (cron, systemd, or a daemon loop)
7. Your scheduler runs the directives — AIMFP has no timer and is not running at 3am
8. The runner records each execution through aimfp.hooks; logs land in .aimfp-project/logs/, statistics in user_directives.db
9. Next session, AIMFP reports what happened and what looks broken

Note: User directive files stay in the user's project. .aimfp-project/ is AI-managed metadata.

Directive Configuration Schema

Triggers and actions have an enforced grammar. A non-conforming config is refused at insert — a bad action fails loudly at runtime, but a bad schedule fails silently by never firing, so both are checked before they are stored.

Triggers — time triggers are a discriminated union on kind:

{"kind": "interval", "seconds": 900}{"kind": "daily",    "at": "17:00", "timezone": "America/New_York"}{"kind": "weekly",   "at": "17:00", "weekdays": ["mon","wed","fri"]}{"kind": "monthly",  "at": "09:00", "days": [1, 15]}
{"event": "stove_on", "source": "home_assistant"}              // event{"expression": "cpu > 90", "evaluate_every_seconds": 60}       // condition (fires once per rise){"expression": "cpu > 90", "repeat": "while_true"}             // condition (fires every true check){}                                                             // manual

at is 24-hour HH:MM, zero-padded. timezone is an IANA name, and is not valid on an interval — an interval counts elapsed seconds. Unknown keys are errors, because a typo'd schedule key would otherwise validate cleanly and then never fire.

Actions — an envelope per action_type:

{"endpoint": "/lights/off", "api": "homeassistant", "method": "POST"}  // api_call{"script": "scripts/backup.sh", "args": ["--full"]}                    // script_execution{"command": "systemctl restart nginx"}                                 // command{"function": "handlers.lights.turn_off"}                               // function_call{"message": "Backup finished", "channel": "ops"}                       // notification

For function_call and command, AIMFP validates the envelope only — the function or executable exists in your generated project, so whether it actually resolves is yours to check.

AIMFP also owns the schedule arithmetic: aimfp.hooks.schedule.next_fire_time computes the next run (wall-clock-correct across DST boundaries), and record_execution_end advances next_scheduled_time automatically after every run. Your runner never needs to hand-roll calendar maths — that is how two projects end up disagreeing about when 17:00 is.

One consequence to know if you build a UI: next_scheduled_time is stored as naive local time, because is_due compares it against a naive local now. A schedule declared as 17:00 in America/New_York is stored as 14:00 on a machine in Los Angeles. Nothing is lost — the declared zone stays in trigger_config — but the column is a comparison value, not a rendering. Show a user their schedule from trigger_config, never from next_scheduled_time, or they will "fix" a config that was already correct.

AIMFP never evaluates a condition, but the grammar says how often a held one fires. By default a condition fires once when it becomes true and re-arms when an evaluation finds it false; "repeat": "while_true" fires on every true evaluation. The runner keeps that latch with set_condition_state, so a restart does not re-fire a condition that is still true. A time slot the runner reaches too late is skipped, never caught up, and recorded with record_directive_skip. A skip counts as neither a run nor an error, and a directive that keeps skipping is reported skipping rather than a false ok.

The AI validates configs with the validate_trigger_config and validate_action_config tools, which return the expected shape alongside any errors. The full spec is available as data from the trigger_config_grammar() and action_config_grammar() hooks.

How It Works

1. AIMFP MCP Gateway Pattern

The aimfp_run command serves as a gateway and reminder, not an executor. It tells the AI that AIMFP directives should be applied.

Every aimfp_run call returns:

{  "success": true,  "message": "AIMFP MCP available",  "guidance": {    "directive_access": "Directive names cached from session bundle. Query specific directives by name when needed.",    "when_to_use": "Use AIMFP directives when coding or when project management action/reaction is needed.",    "assumption": "Always assume AIMFP applies unless user explicitly rejects it."  }}

The MCP server exposes CRUD helper functions for all database operations — tracking files, functions, tasks, project state, user preferences, and (for Use Case 2) automation directives. AI discovers available helpers from the database at runtime.

AI Decision Flow:

1. User prefixes request with aimfp run (or AI assumes it)
2. AI calls aimfp_run tool → receives guidance
3. AI evaluates: Is this coding or project management?
4. If yes: Check if directives are in memory
   • No directives? → Directive names cached from session bundle; query by name when needed
   • Has directives? → Apply appropriate ones
5. If no: Respond without directives

2. Command Flow Example

User: "Help me build a calculator"

AI Processing:

1. Calls aimfp_run(is_new_session=true) → receives session bundle (directive names, project status, settings, supportive context)
2. Checks project state from bundle: .aimfp-project/ missing → project not initialized
3. Calls aimfp_init helper (Phase 1: mechanical setup)
   • Programmatically creates .aimfp-project/ directory, databases, blueprint template
4. Enters Phase 2 (intelligent population): detects language/tools, discusses project with user
5. Routes to project_discovery: collaborates with user on blueprint, themes, flows, completion path, milestones
6. Discovery complete → routes to aimfp_status → project_progression → first task created
7. Work begins

3. Self-Assessment Framework

Before acting, AI performs self-assessment using questions provided with directives:

Core Questions:

1. Does this involve coding or project state changes?

   • Coding and project management are unified — writing code triggers file tracking, DB updates, and task progress automatically
   • FP baseline is always active as the mandatory coding style
   • Project directives handle the tracking side (file writes, DB updates, task management)

2. Do I have directives in memory?

   • No: Load from session bundle via aimfp_run(is_new_session=true)
   • Yes: Proceed with cached directives

3. Which directives apply?

   • FP baseline: Mandatory coding style (pure functions, immutability, no OOP) — applied naturally, not as directive calls
   • FP directives: Reference documentation only — consulted when uncertain about complex scenarios
   • Project directives: File writes, DB updates, task management

4. Action-reaction needed?

   • Code write → project_file_write directive → DB update
   • File edit → DB sync
   • Discussion with decision → DB update

Example Flow (Coding Task):

User: "Write multiply_matrices function"AI thinks:  ✓ FP baseline applies (write pure, immutable, no OOP)  ✓ Project tracking applies (project_file_write directive)  ✓ I have directives in memory
AI executes:  1. Write function following FP baseline naturally  2. Apply project_file_write directive  3. Update project.db (files, functions, interactions)

4. Directive Execution

Directives follow a trunk → branches → fallback pattern:

{  "trunk": "analyze_function",  "branches": [    {"if": "pure_function", "then": "mark_compliant"},    {"if": "mutation_detected", "then": "refactor_to_pure"},    {"if": "low_confidence", "then": "prompt_user"},    {"fallback": "prompt_user"}  ]}

Getting Started

Prerequisites

• Python 3.11+ (required for type hint syntax used throughout; developed and tested through 3.14)
• Linux — see Platform Support below

Platform Support

AIMFP is developed and tested on Linux only. There is no platform-conditional code anywhere in the codebase, so Windows and macOS support is unbuilt rather than untested — it has never been attempted, and making it work would be real porting effort, not a compatibility pass.

The core of the server is portable in principle (pure-stdlib JSON-RPC over stdio, SQLite, pathlib), so parts of it may happen to run elsewhere. Nothing about that is verified or supported. If you run AIMFP on another platform, you're on your own — bug reports are welcome, but Linux is the only environment the project targets.

Installation

AIMFP installs three ways. On Claude Code, use the plugin (Method 1) — two commands, no manual setup. pip/manual are for other MCP clients or advanced setups.

Method 1: Claude Code Plugin (Recommended for Claude Code)

Prerequisite: uv on your PATH (a single self-contained binary). The plugin launches the MCP server with uvx, which fetches the published aimfp package from PyPI, resolves its watchdog dependency, and provisions a compatible Python (3.11+) automatically — so you don't need a matching system Python or a separate pip install.

/plugin marketplace add aryanduntley/aimfp/plugin install aimfp@aimfp

That's it — no pip install, no claude mcp add, no config files. The plugin provides:

• the MCP server, launched via uvx aimfp==<plugin version> — each plugin release runs the matching PyPI release, which uvx caches, so the server starts offline after the first launch and moves forward when you update the plugin
• slash commands: /aimfp:run, /aimfp:status, /aimfp:init, /aimfp:end
• a tiny aimfp-mode setup skill + a SessionStart hook — these are setup-only: they prompt the AI to install the system prompt (via the get_system_prompt tool) and start a session. They are deliberately not a copy of the behavioral rules (that would duplicate the system prompt into every session). Your first move after install is still to ask the AI to add the AIMFP system prompt — see Add the System Prompt below; for the plugin the AI does it for you, nothing to download or paste.
• session hooks for write-tracking nudges and an end-of-session audit guard

Updates: /plugin marketplace update aimfp. No Anthropic account, approval, or central registry is involved — the marketplace is just this public GitHub repo.

Tool pre-approval still applies to the plugin's MCP tools (named mcp__plugin_aimfp_aimfp__*; the allowlist covers both naming schemes) — see Pre-Approve All AIMFP Tools below.

Use one route per project. If a project's .mcp.json already registers an aimfp server (Method 2/3), installing the plugin starts a second AIMFP server and every tool appears twice (mcp__aimfp__* and mcp__plugin_aimfp_aimfp__*). Remove the .mcp.json entry, or don't enable the plugin in that project.

Method 2: pip install

pip install aimfp

This installs the MCP server and makes the aimfp command available. AIMFP requires only one external dependency (watchdog for filesystem monitoring) — the JSON-RPC server itself is pure Python stdlib.

Method 3: Manual Install (GitHub Download)

1. Download the repository (zip download or git clone)
2. Locate the src/aimfp/ folder — this is the complete MCP server package
3. Copy the aimfp/ folder to wherever you keep MCP servers:
   bash

   # Example: copy to your MCP servers directorycp -r src/aimfp/ ~/mcp-servers/aimfp/

The only runtime dependency (watchdog) is installed automatically. The aimfp/ folder contains everything else the server needs: helper functions, directives, database schemas, and the pre-populated aimfp_core.db.

Add the System Prompt (Your First Move)

Do this before anything else. The system prompt is AIMFP's record-keeping backbone — it's what tells the AI to call aimfp_run() first and run the project through AIMFP. Without it the MCP tools are just passive functions that never get called; the AI won't even know to start init or discovery.

Recommended — works for every install method, including the plugin (nothing to download or paste). Just ask the AI:

Add the AIMFP system prompt

The AI calls the get_system_prompt tool, which returns the prompt plus placement guidance, and writes it for you:

• It goes in CLAUDE.md at your project root (Claude Code) or Settings → Custom Instructions (Claude Desktop).
• AIMFP content is placed first — it's the project's backbone and must be the highest-priority instruction.
• If you already have an extensive CLAUDE.md / instructions file, the AI will not blindly prepend — it reviews it with you and discusses consolidating/optimizing first. Existing content is never discarded.
• Once it's in place, the AI offers to also set up the tool-permission allowlist — a one-time step that skips per-tool approval prompts. It only pre-approves AIMFP's own CRUD tools, which operate strictly inside .aimfp-project/: no network, no credentials, nothing outside your current project.

Manual alternative (other MCP clients, or if you'd rather paste it yourself) — print it and paste it into your client's system-prompt / custom-instructions field:

python3 -m aimfp --system-prompt

Configure Your AI Client

Register the AIMFP MCP server in your AI client's configuration. The server uses stdio transport — it reads JSON-RPC messages from stdin and writes responses to stdout.

Claude Desktop

Edit claude_desktop_config.json:

{  "mcpServers": {    "aimfp": {      "command": "python3",      "args": ["-m", "aimfp"],      "env": {}    }  }}

If you used Method 3 (manual install), add the parent directory of your aimfp/ folder to PYTHONPATH so Python can find it:

{  "mcpServers": {    "aimfp": {      "command": "python3",      "args": ["-m", "aimfp"],      "env": {        "PYTHONPATH": "/path/to/parent-of-aimfp-folder"      }    }  }}

For example, if you copied aimfp/ to ~/mcp-servers/aimfp/, set PYTHONPATH to ~/mcp-servers.

If you installed into a virtual environment, use the full path to the venv's Python so Claude Desktop uses the correct interpreter:

{  "mcpServers": {    "aimfp": {      "command": "/path/to/venv/bin/python3",      "args": ["-m", "aimfp"],      "env": {}    }  }}

Claude Code

Use claude mcp add to register the server. Run the command from within the project folder you want AIMFP to manage.

Choose a scope first. AIMFP enforces strict functional programming and actively rejects OOP codebases. If you work on other projects that use OOP or don't need AIMFP, avoid --scope user — it enables the server in every project you open.

pip install (system-wide):

claude mcp add --transport stdio --scope project aimfp -- python3 -m aimfp

pip install (virtual environment) — use the venv's Python path:

claude mcp add --transport stdio --scope project aimfp -- /path/to/venv/bin/python3 -m aimfp

A bare python3 resolves to the system Python, which won't have the package. Use the full path to the venv's interpreter so the MCP server subprocess finds the installed package.

Manual install or running from source — set PYTHONPATH to the parent of the aimfp/ folder:

claude mcp add --transport stdio --scope project --env PYTHONPATH=/path/to/parent-of-aimfp aimfp -- python3 -m aimfp

Quick reference:

Note: All flags (--transport, --scope, --env) must come before the server name. The -- separates the name from the command. Verify the server is connected with /mcp inside Claude Code.

Claude Code: Pre-Approve All AIMFP Tools (Optional)

What these tools actually do: AIMFP operates entirely inside the project you point it at. It never reads, writes, or reaches anything outside that project directory — and within it, all of its own state lives in a single .aimfp-project/ folder. The bulk of the 250+ tools are plain CRUD operations against the local SQLite database at .aimfp-project/project.db (tracking tasks, milestones, modules, and project state) plus read-only lookups of the bundled project directives. There are no network calls, no credentials, no access to your wider filesystem.

Claude Code's design prompts you to approve each MCP tool the first time it's called. With 250+ tools that's one one-time prompt per tool — tedious, but every prompt is just Claude Code confirming a local database or directive call. To skip the prompts, drop a Claude Code allowlist file (.claude/settings.local.json) into your project. This is a static Claude Code settings file that simply tells Claude Code "these tool names are pre-approved" — it is not downloaded or executed by AIMFP.

Just ask the AI to set it up:

"Set up AIMFP permissions for Claude Code"

The AI calls the get_claude_permissions tool, which returns the complete, current allowlist (generated live from the tool registry, so it can never be out of date or incomplete). The AI then writes — or, if you already have a .claude/settings.local.json, merges into — that file. Merging preserves every non-AIMFP permission and any other settings you already have; only the AIMFP entries are refreshed. Each tool is allowlisted under both names Claude Code can give it — mcp__aimfp__* (a .mcp.json server) and mcp__plugin_aimfp_aimfp__* (the plugin) — so the file is right for either install route. AIMFP itself never writes the file (it never touches anything outside the project) — the AI does, with its normal file-write confirmation. The first call to get_claude_permissions is the one prompt you approve to eliminate all the rest.

The two MCP-autostart keys (enableAllProjectMcpServers and enabledMcpjsonServers) are included automatically so the AIMFP server starts itself when you have a .mcp.json in the project.

Note: A pre-built copy of this allowlist also ships at documentation/settings.local.json if you'd rather review it directly or copy it in by hand (cp documentation/settings.local.json /path/to/your/project/.claude/). The get_claude_permissions tool is the recommended path since it's always in sync with the installed version.

Other MCP Clients

The server uses stdio transport. Point your client at python3 -m aimfp (or the full path to your venv's Python). For manual installs, ensure PYTHONPATH includes the parent directory of the aimfp/ folder. No API keys or authentication required.

How It Works

The server communicates over stdio using the Model Context Protocol. It resolves aimfp_core.db (the directive database) relative to its own installation — no environment variables needed.

Once connected, the AI calls aimfp_run() on every interaction (guided by the system prompt). Project state is stored in .aimfp-project/ in your working directory, created automatically when you initialize a project.

Project Initialization

Tell your AI assistant:

"Initialize AIMFP for my project"

The AI calls aimfp_init which creates an .aimfp-project/ folder in your project root containing databases for project state tracking, user preferences, and a ProjectBlueprint document. You don't need to interact with these files — the MCP server manages them automatically.

Directives System

FP Baseline vs FP Directives

FP Baseline (Always Active):

• Core functional programming rules AI follows naturally when writing code
• Pure functions, immutability, no OOP, explicit error handling
• Non-negotiable - all code must be FP-compliant

FP Directives (Reference Documentation):

• Detailed guidance for complex scenarios and edge cases
• Consulted only when AI is uncertain about implementation
• Categories: Purity, Composition, Error Handling, OOP Elimination, Optimization
• Examples: fp_purity, fp_monadic_composition, fp_result_types, fp_wrapper_generation

Project Directives

Manage project lifecycle:

User Preference Directives

Manage AI behavior customization and learning:

User-Defined Directives

FOR USE CASE 2 ONLY: Automation projects where AIMFP generates and manages the codebase:

Use Cases (Automation Projects):

• Home Automation: "At 5pm turn off living room lights", "If stove on > 20 min, turn off"
• Cloud Infrastructure: "Scale EC2 when CPU > 80%", "Backup RDS nightly at 1am"
• Custom Workflows: "Every Monday generate report", "Process uploaded files automatically"

Key Architecture:

• User writes directive definitions (YAML/JSON/TXT)
• AIMFP generates the entire automation codebase (src/, tests/, etc.)
• AIMFP manages the generated code like any software project (tasks, files, functions)
• AIMFP never executes a directive and never supervises the runner — the generated code owns the timer and the work; AIMFP answers "what is due" and records "what happened" through the aimfp.hooks library API
• A dead runner is detected without any process inspection: the runner declares when it will next fire, and silence past that deadline is the evidence
• Project.db tracks the generated code; user_directives.db tracks directive state

Key Features:

• Write directives in YAML, JSON, or plain text
• AI validates through interactive Q&A
• AI generates complete FP-compliant Python modules in src/
• AI creates tests for generated code
• Execution driven by your own scheduler, recorded through the AIMFP hook library
• File-based logging (30-day execution logs, 90-day error logs) written by AIMFP, so every automation project's records have the same shape
• Health reporting at session start: overdue, skipping, degraded, never-run, or in error
• Dependency management with user confirmation

Example Directive Definition:

# directives/home_automation.yaml (user creates this in their project)directives:  - name: turn_off_lights_5pm    trigger:      type: time      time: "17:00"      timezone: America/New_York    action:      type: api_call      api: homeassistant      endpoint: /services/light/turn_off      params:        entity_id: group.living_room_lights

User tells AI: "Parse my directive file at directives/home_automation.yaml"

AI Generates (in src/lights_controller.py):

# Auto-generated from home_automation.yamlfrom typing import Resultfrom homeassistant_client import HomeAssistant
def turn_off_living_room_lights(ha_client: HomeAssistant) -> Result[None, str]:    """Turn off all lights in living room group."""    # FP-compliant implementation    ...

See the directive MD files in src/aimfp/reference/directives/ for complete workflow documentation.

Git Integration

FP-powered Git collaboration for multi-user and multi-AI development:

Key Features:

• Multi-user collaboration: Multiple developers and AI instances work simultaneously
• FP-powered conflict resolution: Uses purity levels, dependencies, and test results to auto-resolve conflicts
• Branch naming: aimfp-alice-001, aimfp-bob-002, aimfp-ai-claude-001
• Auto-resolution: High-confidence conflicts (>0.8) resolved automatically using FP purity rules
• External change detection: Compares Git HEAD with stored hash to detect changes made outside AIMFP
• Simplified schema: No separate git_state table - Git commands are fast enough (~1ms)

Why AIMFP + Git is Superior to OOP + Git:

Example Conflict Resolution:

Alice and Bob both modify calculate_total():
Alice's version:- Purity: ✅ Pure function- Tests: 10/10 passing
Bob's version:- Purity: ✅ Pure function- Tests: 12/12 passing (includes edge cases)
AI Recommendation: Keep Bob's version (confidence: 85%)Reason: More comprehensive tests, still pure

See the Git directive MD files in src/aimfp/reference/directives/ for complete multi-user collaboration workflows.

Project Lifecycle

graph TD    A[aimfp_init] --> B[project_discovery]    B --> C[project_progression]    C --> D[project_file_write]    D --> E[project_update_db]    E --> F{All Items Done?}    F -->|No| D    F -->|Yes| G[project_task_complete]    G --> H{All Tasks Done?}    H -->|No| C    H -->|Yes| I[project_milestone_complete]    I --> J{All Milestones Done?}    J -->|No| C    J -->|Yes| K[project_completion_check]    K --> L[project_archive]

Example Workflow

Create Project

User: "Help me build a matrix calculator"
AI → aimfp_run(is_new_session=true)    → Receives: session bundle (directive names, status, settings, supportive context)    → Checks project state: .aimfp-project/ missing → not initialized    → Calls aimfp_init helper (Phase 1: mechanical setup)        → Creates .aimfp-project/ directory        → Creates project.db, user_preferences.db        → Copies ProjectBlueprint template    → Enters Phase 2 (intelligent population)        → Detects language, build tool, source directory        → Discusses with user: project name, purpose, goals        → Updates infrastructure table and blueprint    → Routes to project_discovery        → Collaborates with user on themes, flows, completion path, milestones
AI: ✅ Project initialized: MatrixCalculator    Next: Let's define the project shape — themes, completion path, and first milestones.

Write Function

User: "Write multiply_matrices function"
AI → aimfp_run() (continuation — lightweight guidance)    → Evaluates: coding task + project tracking needed    → Has directives cached from session bundle    → Writes FP-compliant function naturally (pure, immutable, no OOP)    → Follows project_file_write directive workflow:        → Reserves file and function IDs in project.db        → Writes src/matrix.py with embedded IDs        → Finalizes file and function entries        → Updates project.db:           • files table (new entry)           • functions table (multiply_matrices)           • interactions table (dependencies)
AI: ✅ Function written: multiply_matrices    File: src/matrix.py:15    DB Updated: ✓ files, functions, interactions    Dependencies: validate_dimensions

Check Compliance

User: "Check project compliance"
AI → aimfp_run() (continuation — lightweight guidance)    → project_compliance_check directive (opt-in tracking feature)    → Queries all functions from project.db    → Reviews each function against FP baseline    → Generates analytics report
AI: 📊 Compliance Report:    Total functions: 12    Compliant: 10    Violations: 2      - calculate_discount (uses global config)      - process_order (hidden mutation)

Customize AI Behavior

User: "Always add docstrings to functions"
AI → aimfp_run() (continuation — lightweight guidance)    → user_preferences_update directive        → Calls find_directive_by_intent helper        → Searches directives by name/description/keywords        → Finds match: project_file_write        → Confirms with user: "Apply to file writing?"        → Updates user_preferences.db
AI: ✅ Preference saved: project_file_write    Key: always_add_docstrings    Value: true
    Next file write will include docstrings automatically.

AI Learns from Corrections

User: "Actually, use guard clauses instead of nested if statements"
AI → user_preferences_learn directive (detects correction)    → Logs interaction to ai_interaction_log    → Infers preference: prefer_guard_clauses = true    → Prompts user: "Should I remember this preference?"
User: "Yes"
AI: ✅ Preference learned: project_file_write    Key: prefer_guard_clauses    Value: true
    I'll use guard clauses in future functions.

Documentation

Directive Reference

All directive documentation is shipped with the package at src/aimfp/reference/directives/ — 131 MD files covering every directive. Each file includes: purpose, when to apply, complete workflows (trunk → branches), compliant/non-compliant examples, edge cases, related directives, helper functions used, and database operations.

Database Schemas

Schema SQL files are in the package at src/aimfp/database/schemas/:

• aimfp_core.sql — Global read-only database (directives, helpers, flows)
• project.sql — Per-project mutable database (files, functions, tasks, milestones)
• user_preferences.sql — Per-project user customization database
• user_directives.sql — Per-project automation directives (Use Case 2 only)

Development

Dev Directory (dev/)

The dev/ directory contains the source of truth for directive and helper function definitions. These JSON files are the canonical definitions that get imported into aimfp_core.db before release.

dev/├── directives-json/              # Directive definitions (source of truth)│   ├── directives-fp-core.json   # Core FP directives│   ├── directives-fp-aux.json    # Auxiliary FP directives│   ├── directives-project.json   # Project lifecycle directives│   ├── directives-user-pref.json # User preference directives│   ├── directives-user-system.json # User automation directives│   ├── directives-git.json       # Git collaboration directives│   ├── directive_flow_fp.json    # FP directive flow transitions│   ├── directive_flow_project.json # Project directive flow transitions│   └── directive_flow_user_preferences.json # User pref flow transitions├── helpers-json/                 # Helper function definitions (source of truth)│   ├── helpers-core.json         # Core/directive helpers│   ├── helpers-orchestrators.json # Entry point and status helpers│   ├── helpers-project-*.json    # Project management helpers (9 numbered files)│   ├── helpers-project-modules.json    # Module helpers│   ├── helpers-project-task-files.json # Task-file link helpers│   ├── helpers-catalog.json      # Existing-codebase adoption helpers│   ├── helpers-changeset.json    # Semantic changeset merge helpers│   ├── helpers-backup.json       # Backup helpers│   ├── helpers-notices.json      # One-time release notice helpers│   ├── helpers-settings.json     # User preference helpers│   ├── helpers-user-custom.json  # User directive helpers│   ├── helpers-git.json          # Git operation helpers│   └── helpers-index.json        # Shared/global helpers├── notices-json/notices.json     # One-time release notices (synced into aimfp_core.db)├── sync-directives.py            # Imports JSON → aimfp_core.db├── bump-version.py               # Syncs the package version across the files below└── logs/                         # Development logs

Dev workflow: Modify JSON files in dev/ → delete src/aimfp/database/aimfp_core.db and run sync-directives.py to rebuild it → test → release. End users only interact with the pre-populated aimfp_core.db, never the JSON files directly.

Version Locations

All four must be kept in sync when bumping versions (python3 dev/bump-version.py does this):

Design Philosophy

Immutable Rules, Evolving Projects

• aimfp_core.db is the rulebook (read-only, global, version-controlled)
• project.db is the workspace (read-write, per-project, runtime state)
• user_preferences.db is the customization layer (read-write, per-project, AI behavior)
• Directives define the boundaries; AI operates freely within them

Database-Driven Context

Traditional AI assistants lack persistent memory. AIMFP solves this:

-- AI remembers everything across sessions via helper toolsAI calls: get_project_status("detailed")→ Returns: project metadata, active milestone, current task, recent notes, tracked files and functions
-- No source code reparsing required. Instant context retrieval.

Finite by Design

Every AIMFP project has a completion path:

Setup → Core Development → Testing → Documentation → Finalization

Once project_completion_check passes, the project is done. No endless feature creep.

Features

• Comprehensive MCP tools — Full CRUD for 4 SQLite databases, covering project management, FP directives, user preferences, and custom automation directives
• Pure functional enforcement — AI writes FP-compliant code by default (pure functions, immutability, no OOP)
• Database-driven persistent memory — Project state survives across sessions; no context loss
• Directive-based workflows — Deterministic trunk → branches → fallback execution patterns
• Finite completion paths — Projects have defined stages, milestones, and tasks; work converges toward completion
• Two use cases — Regular software development (Use Case 1) or custom directive automation (Use Case 2)
• User preference learning — AI adapts to coding style via per-directive key-value overrides
• Git integration — FP-powered branch management and conflict resolution
• Minimal dependencies — One runtime package (watchdog), custom JSON-RPC server uses stdlib only
• Cost-conscious design — All tracking/analytics features disabled by default

Token Overhead

AIMFP's project management adds token overhead to each session. Database operations (tracking files, updating tasks, managing milestones) consume tokens beyond the code-writing itself. We work to minimize this — continuation calls are lightweight (~2k tokens), batch helpers reduce round-trips, and all optional tracking is off by default.

The payoff comes with project scale. Without AIMFP, every new session requires the AI to re-scan directories, re-read source files, and rebuild context from scratch — costs that grow with codebase size. With AIMFP, the database provides instant structured context: the AI knows exactly which files exist, what functions they contain, where work left off, and what comes next. For projects beyond a handful of files, the up-front tracking cost is recovered many times over through eliminated re-discovery work.

Usage Examples

Example 1: Project Initialization

User prompt: "Help me build a calculator"

Tool calls:

1. aimfp_run(is_new_session=true)   → Returns: session bundle (directive names, settings, project status, supportive context)   → Status shows: .aimfp-project/ missing — not initialized
2. aimfp_init(project_root="/home/user/calculator")   → Returns: { success: true, message: "Project initialized" }

What happens:

• AI detects no project exists and automatically calls aimfp_init helper
• Creates .aimfp-project/ directory with project.db, user_preferences.db, and ProjectBlueprint.md
• Registers project metadata (name, root path, infrastructure)
• Inserts default user settings and backup configuration
• AI enters Phase 2: discusses project shape with user (infrastructure, purpose, goals)
• Routes to project_discovery for collaborative planning (themes, flows, completion path, milestones)

Example 2: Writing FP-Compliant Code

User prompt: "Write a multiply_matrices function"

Tool calls:

1. aimfp_run(is_new_session=false)   → Returns: lightweight guidance
2. reserve_file(name="matrix_operations", path="src/matrix_operations.py", language="Python")   → Returns: { success: true, id: 42 }
3. reserve_function(name="multiply_matrices", file_id=42, purpose="Multiply two matrices")   → Returns: { success: true, id: 99 }
   (AI writes FP-compliant code to src/matrix_operations_id_42.py)
4. finalize_file(file_id=42, path="src/matrix_operations_id_42.py", ...)   → Returns: { success: true }
5. finalize_function(function_id=99, name="multiply_matrices_id_99", file_id=42, purpose="...", parameters=[...], returns={...})   → Returns: { success: true }

What happens:

• File and function are reserved in project.db before writing (IDs embedded in names)
• AI writes pure functional code following FP baseline (no OOP, no mutations, explicit parameters)
• File and function are finalized with purpose, parameters, and return metadata
• Project database now tracks the code structure for instant retrieval in future sessions

Example 3: Resuming Work / Checking Status

User prompt: "Where are we?" or "Continue working"

Tool calls:

1. aimfp_run(is_new_session=false)   → Returns: guidance + common starting points
   (AI answers from cached context — no additional DB calls needed    unless context is stale, in which case:)
2. aimfp_status(type="detailed")   → Returns: project metadata, active milestone, current task, recent notes, warnings

What happens:

• AI presents current milestone, active task, and next steps
• If project_continue_on_start=true, AI automatically picks up the next task
• No source code reparsing needed — everything is indexed in project.db

Privacy Policy

AIMFP runs entirely on your local machine. No data ever leaves your computer.

• No network requests: The MCP server makes zero network calls. All operations are local SQLite database reads/writes and filesystem operations.
• No telemetry: No usage analytics, crash reports, or diagnostic data is collected or transmitted.
• No third-party services: AIMFP does not connect to any external APIs, cloud services, or remote servers.
• Local data only: All project data (databases, preferences, directives) is stored in your project's .aimfp-project/ directory. You own and control all data.
• No Claude memory access: AIMFP does not access or read Claude's conversation history, memory, or any other client-side data.

If you have questions about data handling, please open an issue on our GitHub repository.

Support

• Issues: GitHub Issues
• Repository: github.com/aryanduntley/aimfp

Contributing

AIMFP is an open standard for AI-optimized programming. Contributions welcome:

1. New FP directives — Language-specific or advanced patterns
2. Helper functions — Database, file, Git, or FP utilities
3. Templates — ADT boilerplate, error handling patterns
4. Documentation — Examples, tutorials, case studies

By submitting a contribution (pull request, patch, or other material), you agree that it may be distributed under the PolyForm Noncommercial License 1.0.0 and under any commercial license the copyright holder grants for AIMFP, and you confirm you have the right to make that grant.

License

AIMFP is licensed under the PolyForm Noncommercial License 1.0.0. See LICENSE for the full terms.

Free for noncommercial use: personal study, research, experimentation, hobby projects, and use by charitable organizations, educational institutions, public research organizations, and government institutions.

Commercial use requires written approval. This covers any for-profit use, including use by or for a business, use in paid or client work, selling or redistributing AIMFP or a derivative of it, and offering it as part of a paid product or service. To request a commercial license, contact Aryan Duntley at [email protected].

Versions up to and including 1.66.1 were released under the MIT License, and copies of those versions remain under MIT. Every later version is licensed only under the PolyForm Noncommercial License 1.0.0.

Summary

AIMFP transforms AI from a "code generator" into a structured, customizable project collaborator.

It combines:

• Pure functional programming for deterministic, composable code
• Four-database architecture for immutable rules, runtime state, user customization, and optional automation
• Directive-based workflows for consistent, automated compliance
• User preference learning for AI that adapts to your coding style
• Finite completion paths for goal-oriented development
• Cost-conscious design with opt-in tracking features

The result: AI-maintained codebases that are predictable, traceable, customizable, and maintainable across sessions, teams, and even different AI assistants.

Built for the age of AI-native development.