Boyce

Boyce: Semantic Protocol & Safety Layer for Agentic Database Workflows

The semantic safety layer for agentic database workflows. Boyce connects LLMs to live database context with built-in safety rails.

Named for Raymond F. Boyce, co-inventor of SQL (1974) and co-author of Boyce-Codd Normal Form (BCNF).

AI agents querying databases without proper context generate unreliable SQL — working from incomplete schemas, inferring column names, guessing join paths. Boyce gives agents the structured database intelligence they need to generate correct, safe SQL every time — through three interconnected systems:

Why does this matter? → The Null Trap: Your AI Agent's SQL Is Correct. The Answer Is Still Wrong.

Install

Requires Python 3.10+

pip install boyce
# With live Postgres/Redshift adapter (enables EXPLAIN pre-flight + column profiling)pip install "boyce[postgres]"

# uv (recommended)uv pip install boyceuv pip install "boyce[postgres]"

From source:

git clone https://github.com/boyce-io/boyceuv pip install -e "boyce/"

Quickstart

After installing, run boyce init to configure your MCP host automatically:

boyce init

The wizard detects Claude Desktop, Cursor, Claude Code, and JetBrains (DataGrip, IntelliJ, etc.), and writes the correct config block for each.

Developing from source? The repo includes a setup script:

./quickstart.sh   # detects uv or python, installs package, writes .env template

Configure Your MCP Host

The fastest path is boyce init — it detects your MCP host and writes the config automatically:

boyce init

Or configure manually. There are two setup paths depending on your host:

Path 1 — MCP Hosts (No LLM key required)

If you're using Claude Desktop, Cursor, Claude Code, Codex, Cline, Windsurf, JetBrains (DataGrip, IntelliJ), or any MCP-compatible host, you do not need to configure an LLM provider for Boyce. The host's own model handles reasoning — Boyce supplies the schema context and deterministic SQL compiler via get_schema and ask_boyce. Only BOYCE_DB_URL is needed (and even that is optional).

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{  "mcpServers": {    "boyce": {      "command": "boyce",      "env": {        "BOYCE_DB_URL": "postgresql://user:pass@host:5432/db"      }    }  }}

Cursor (.cursor/mcp.json in project root):

{  "mcpServers": {    "boyce": {      "command": "boyce",      "env": {        "BOYCE_DB_URL": "postgresql://user:pass@host:5432/db"      }    }  }}

Path 2 — With Boyce's Built-in NL→SQL

If you're using the CLI (boyce ask), HTTP API, or a non-MCP client (e.g., the VS Code extension), configure Boyce's internal query planner with your LLM provider:

{  "mcpServers": {    "boyce": {      "command": "boyce",      "env": {        "BOYCE_PROVIDER": "anthropic",        "BOYCE_MODEL": "claude-sonnet-4-6",        "ANTHROPIC_API_KEY": "sk-ant-...",        "BOYCE_DB_URL": "postgresql://user:pass@host:5432/db"      }    }  }}

Boyce supports any LLM provider available through LiteLLM: Anthropic, OpenAI, Ollama (local), vLLM (local), Azure, Bedrock, Vertex, Mistral, and more.

BOYCE_DB_URL is optional on both paths. Without it, Boyce runs in schema-only mode — SQL generation still works; EXPLAIN pre-flight and live query tools return "status": "unchecked".

Environment Variables

MCP Tools

Architecture

SemanticSnapshot (JSON)        │        ▼  ingest_source ┌─────────────────────────────────────────────┐ │          SemanticGraph (NetworkX)            │  ← in-memory, loaded per session │  nodes = entities (tables/views/dbt models) │ │  edges = joins  (weighted by confidence)    │ └─────────────────────────────────────────────┘        │                         │        ▼  ask_boyce              ▼  (internal)  QueryPlanner                 Dijkstra  (LiteLLM)                    join resolver  NL → StructuredFilter             │        │                           │        └──────────┬────────────────┘                   ▼           kernel.process_request()          ← ZERO LLM HERE           SQLBuilder (dialect-aware)                   │                   ▼           EXPLAIN pre-flight                ← Query Verification           (PostgresAdapter)                   │                   ▼            SQL + validation result

Dialect support: redshift, postgres, duckdb, bigquery

Redshift safety rails (safety.py): Automatic linting for LATERAL, JSONB, REGEXP_COUNT, lookahead regex patterns, and numeric cast rewrites for Redshift 1.0 (PG 8.0.2).

Scan CLI

# Scan a single fileboyce scan demo/magic_moment/manifest.json
# Scan a directory (auto-detects all parseable sources)boyce scan ./my-project/ -v
# Save snapshots for MCP server useboyce scan ./my-project/ --save

10 parsers: dbt manifest, dbt project, LookML, SQLite, DDL, CSV, Parquet, Django, SQLAlchemy, Prisma.

Verify the Install

# Unit tests — no DB required, runs in ~4 secondspython boyce/tests/verify_eyes.py
# Expected output:# Ran 15 tests in 3.5s# OK# ✅  All checks passed.

SemanticSnapshot Format

The ingest_source tool accepts a SemanticSnapshot JSON dict. Minimal example:

{  "snapshot_id": "<sha256>",  "source_system": "dbt",  "entities": {    "entity:orders": {      "id": "entity:orders",      "name": "orders",      "schema": "public",      "fields": ["field:orders:order_id", "field:orders:revenue"]    }  },  "fields": {    "field:orders:order_id": {      "id": "field:orders:order_id",      "entity_id": "entity:orders",      "name": "order_id",      "field_type": "ID",      "data_type": "INTEGER"    }  },  "joins": []}

See boyce/tests/live_fire/mock_snapshot.json for a complete field/entity example.

Project Layout

boyce/                          ← PRIMARY — headless FastMCP server + pip package├── boyce/│   ├── server.py               ← MCP entry point (8 tools)│   ├── kernel.py               ← Deterministic SQL kernel│   ├── graph.py                ← SemanticGraph (NetworkX)│   ├── safety.py               ← Redshift compatibility rails│   ├── types.py                ← Protocol contract (Pydantic)│   ├── scan.py                 ← Scan CLI (boyce scan)│   ├── connections.py          ← DSN persistence (ConnectionStore)│   ├── doctor.py               ← Environment diagnostics (boyce doctor)│   ├── sql/                    ← SQLBuilder, dialect layer, join resolver│   ├── parsers/                ← 10 parsers (dbt, lookml, ddl, sqlite, csv, etc.)│   ├── planner/                ← QueryPlanner (LiteLLM → StructuredFilter)│   └── adapters/               ← PostgresAdapter (Eyes)└── tests/    ├── verify_eyes.py          ← 15-test suite, no DB required    ├── test_parsers.py         ← Parser tests (all 10 parsers)    ├── test_scan.py            ← Scan CLI tests    └── live_fire/              ← Docker Compose integration tests

Status

Support

• Troubleshooting guide: docs/troubleshooting.md
• Local LLM setup (Ollama/vLLM): docs/local-llm-setup.md
• Bug reports: GitHub Issues
• Setup help: GitHub Issues
• Email: [email protected] — for issues involving credentials or sensitive config

Copyright 2026 Convergent Methods, LLC. MIT License.