Rails AI Context

[rails-ai-context: ground truth for your coding agent]

Your coding agent guesses your Rails app. Let it ask.

[Works with] [Claude Code] [Cursor] [GitHub Copilot] [OpenCode] [Codex CLI] [Any terminal]

[Gem version] [Downloads] [CI status] [Ruby 3.1 to 4.0] [Rails 7.0 to 8.1] [MIT license] [Listed in the MCP Registry]

Website · Quick start · FAQ · What it does · A closer look · How it works · Tools reference · Contributing

[Questions before you install? Read the FAQ. Do I need MCP support in my AI tool? No. The gem works three ways. 1, MCP server: AI calls tools via the protocol, the best experience. 2, Static files: generated context files (CLAUDE.md, .cursor/rules/, etc.). 3, CLI: the same 35 tools from the terminal, no server needed. More questions the FAQ answers: What does this gem do? Which AI tools are supported? Is this safe for production? Does it work without a database? Do I need to commit the generated files? How does this compare to a hand-written CLAUDE.md?]

[One session with rails-ai-context, each line from a real answer recorded against the sample app in this repository. It asks about a table: the users table's role column is an integer, not null, default 0, and index_users_on_email is unique. Before it calls an association: User has_many articles, dependent destroy, foreign key author_id. Before it writes an action: the before filter authenticate_user! comes from ApplicationController and is skipped on index and show. Before it writes a test: rspec, factories in spec/factories, 3 files, with the traits in_review and published. It traces a method: 4 call sites, among them article_publisher.rb, a service, with the line return false unless @article.publishable?. When the app cannot boot: a STATIC line says the answer serves static analysis and runtime-only data is marked UNAVAILABLE. Over MCP in Claude Code, Cursor, GitHub Copilot, OpenCode and Codex CLI, or as rails-ai-context tool NAME in any terminal.]

Contents

• What it does
• Try it in your Rails app
• A closer look: the schema, the model, the inherited filters, the trace, the app that cannot boot, the six rules
• The tools
• How it works
• What it costs
• Questions you should ask first
• What it will not do
• Install options and configuration
• Troubleshooting
• Development
• Contributing and security
• Learn more

What it does

An assistant writing Rails code knows what an average Rails app looks like. Yours has its own columns, associations, filters and test setup, and the assistant starts every session with none of that. It writes a migration for a column that exists, calls user.posts when the association is user.articles, scaffolds FactoryBot into a fixture suite, and misses the before_action a parent controller adds. You catch it, fix it, re-prompt, and something next to it breaks.

rails-ai-context gives the assistant 35 tools that read the app itself, over MCP from the AI tool or from the command line. Every answer is read from the app and says how sure it is, and most end with a hint for the next call.

Each line comes from the sample app in this repository, a blog with Devise, an admin namespace, a JSON API, a Sidekiq job and a FactoryBot suite. The website records the whole answers every time it is built, by running the gem at that commit: crisnahine.github.io/rails-ai-context.

The install writes the files each AI tool reads, so the assistant knows the tools exist and how to ask: the MCP config, one root context file with six anti-hallucination rules, and rules that load when a model, a controller or a migration is open. The tools are read-only, the SQL tool runs inside a read-only transaction, and files that hold secrets are never served.

Try it in your Rails app

Open your AI tool at the root of your Rails app and paste this. It installs the gem, sets up the AI tool you are in, checks the result and tells you the one step left.

Install the rails-ai-context gem in this Rails app so you can read its schema, models, routes and code through its tools instead of guessing.
1. Run `ruby -v` and `bundle exec rails -v`. rails-ai-context needs Ruby 3.1+ and Rails 7.0+. If either is older, stop and tell me.2. Pick TOOL, the key for the AI tool you are: claude (Claude Code), cursor (Cursor), copilot (GitHub Copilot in VS Code), opencode (OpenCode), codex (Codex CLI). If you are another AI tool, stop and tell me to use the "Another MCP client" prompt instead.3. Run `bundle add rails-ai-context --group development --version '>= 5.35.0'`. If the Gemfile already lists it, run `bundle update rails-ai-context --conservative` instead. Then run `bundle exec rails-ai-context version`: it must print 5.35.0 or newer, the first version with the flags used below. If it is older, the Gemfile pins it: change its rails-ai-context line to `gem "rails-ai-context", ">= 5.35.0", group: :development` and run `bundle update rails-ai-context --conservative`. If Bundler refuses (a frozen or deployment bundle), run `gem install rails-ai-context -v '>= 5.35.0'` instead and use `rails-ai-context init --tools TOOL --defaults` in step 4.4. Run `bundle exec rails generate rails_ai_context:install --tools TOOL --defaults`. If it fails because the app can't boot, run `bundle exec rails-ai-context init --tools TOOL --defaults` instead; that one reads the source files when the app won't start.5. Run `bundle exec rails-ai-context tool onboard` (or `rails-ai-context tool onboard` if you used gem install) and check it describes this app.6. Tell me which files changed, then the "Next step" lines the install printed for TOOL, word for word. I have to do that step before your rails-ai-context MCP tools appear.

Rather type it yourself? In the Gemfile:

bundle add rails-ai-context --group developmentbin/rails generate rails_ai_context:install --tools claude --defaults

Standalone, with no Gemfile change:

gem install rails-ai-contextrails-ai-context init --tools claude --defaults

As a Claude Code plugin, which adds the MCP server and gives each session the six rules, and writes no context files. It runs the installed gem, so install that first, and run it in the app's folder:

gem install rails-ai-contextclaude plugin install rails-ai-context --marketplace crisnahine/rails-ai-context --scope project

Commit the .claude/settings.json it writes; the install generator and init then leave Claude Code's .mcp.json to the plugin (details).

--tools takes claude, cursor, copilot, opencode, codex or all. Without --defaults the generator asks which AI tools you use and what to write. Re-running is safe: it keeps what you have, adds what is missing, and in a CLAUDE.md you wrote it touches only the block between its <!-- BEGIN/END rails-ai-context --> markers.

You need Ruby 3.1 or newer and Rails 7.0 to 8.1, on Linux, macOS or Windows. Optional: brakeman for security_scan, listen for watch, ripgrep for a faster search_code.

Then one step in your AI tool

Every AI tool holds a new project MCP server back until you say yes. The install prints the step for each tool you picked:

Then ask something only your app knows, such as "what columns does the users table have?", and watch it call rails_get_schema instead of guessing. Every other setup has a prompt: no Gemfile change, CLI only, MCP only, a whole team, a folder of apps, nothing committed, MCP over HTTP, Ruby in Docker, another MCP client.

A closer look

Every block below is the gem's own output for the sample app, read without booting it, trimmed: the tier line and the report line that end every answer are cut, and each note names what else is. The website shows each answer whole, recorded at the commit it was built from.

The schema, with the model on it

$ rails-ai-context --app-path site/sample_app tool schema --table users --no-boot## Table: users
**Models:** User**Primary key:** id
| Column | Type | Null | Default ||--------|------|------|---------|| id | bigint | **NO** |  || email | string | **NO** | "" || encrypted_password | string | **NO** | "" || name | string | yes |  || role | integer | **NO** | 0 || active | boolean | **NO** | true || api_token | string | yes |  |
### Indexes- `index_users_on_api_token` on (api_token) (unique)- `index_users_on_email` on (email) (unique)
_Next: `rails_get_model_details(model:"User")` for associations, validations, scopes._

Trimmed to seven of its twelve columns and two of its three indexes. With the app booted, the columns come from the database connection; here they come from db/schema.rb, and db/structure.sql or the migrations would do.

Before it calls an association

$ rails-ai-context --app-path site/sample_app tool model_details --model Article --no-boot# Article [STATIC]
**Table:** `articles`**File:** `app/models/article.rb` (32 lines)**Structure:** class definition(1-1) → includes(2-3) → associations(4-6) → macros(7-8) → validations(9-12) → callbacks(13-15) → scopes(16-18) → instance methods(19-22) → private(23-24) → instance methods(25-32)
## Associations- `belongs_to` **author** (class: User) (inverse_of: :articles) (fk: author_id)- `has_many` **comments** dependent: destroy
## Validations- `presence` on author _(implicit from belongs_to)_- `presence` on title- `length` on title (maximum: 120)- `presence` on slug- `uniqueness` on slug- `presence` on body (if: :published?)
## Enums- `status`: draft(0), in_review(1), published(2) [integer]
## Scopes- `recent` → order(published_at: :desc, created_at: :desc) [STATIC]- `by_author` → where(author: user) [STATIC]- `published` → where(status: :published).where.not(published_at: nil) [STATIC]
## Callbacks- `before_validation`: :assign_slug (on: :create)- `after_commit (on: :update)`: :notify_author (if: :saved_change_to_status?)
## Concerns- **Publishable** - publishable?, publish!

The model is read with the Prism parser. A scope with a literal body is [VERIFIED], one built at runtime is [INFERRED], and in the static tier every record is capped at [STATIC], since no record can claim more than the tier that carries it. The columns, the primary key, the static-tier note on associations, the instance methods and the next-call hint are trimmed out above. The published scope is declared in the Publishable concern, and the Concerns section lists the methods each concern brings.

Before it writes an action

$ rails-ai-context --app-path site/sample_app tool controllers --controller admin/articles --no-boot# Admin::ArticlesController
**Parent:** `Admin::BaseController`**Layout:** `admin` (declared in Admin::BaseController)
## Actions- `index`- `publish`- `show`
## Filters- `before` **authenticate_user!** _(from ApplicationController)_- `before` **require_admin** _(from Admin::BaseController)_- `before` **set_article** (only: show, publish)

Every filter that applies, with where it was inherited from, so the assistant never misses the authenticate_user! two classes up. The schema hints for the models the controller touches, and the next-call hints, follow in the full answer.

Tracing a method

$ rails-ai-context --app-path site/sample_app tool search_code --pattern 'publishable?' --match_type trace --no-boot# Trace: `publishable?`
## Definition**app/models/concerns/publishable.rb:8** in `module Publishable````ruby  def publishable?    in_review? && body.present? && author.active?  end```
## Called from (4 sites)### app/controllers/admin/articles_controller.rb (Controller) → `POST /admin/articles/:id/publish`  13: if @article.publishable?### app/models/concerns/publishable.rb (Model)  13: raise ArgumentError, "#{self.class.name} #{id} is not ready to publish" unless publishable?### app/services/article_publisher.rb (Service)  11: return false unless @article.publishable?### lib/tasks/articles.rake (Other)  5: PublishArticleJob.perform_later(article) if article.publishable?
## Tested by (1 reference)- `spec/models/article_spec.rb` (1 reference)

One call in place of a file-by-file search: the definition, what it calls, its siblings, every caller grouped by file and labelled by kind, with the route a controller caller answers, and the tests. The "Calls internally" and "Sibling methods" sections are trimmed out above.

When the app cannot boot

The app-reading commands try a full boot for live reflection. When the boot fails, from a missing environment variable, an unreachable service or a broken initializer, they answer from the static tier instead of dying: routes from config/routes.rb, the schema from db/schema.rb, db/structure.sql or the migrations, models and controllers from their source. Every answer carries a tier line, a section that needs a booted app says [UNAVAILABLE] with the reason, and --no-boot asks for that tier outright:

_[STATIC] Static mode (static mode requested with --no-boot). Serving static analysis; runtime-only data is marked [UNAVAILABLE]. Run `bundle exec rails-ai-context --app-path site/sample_app doctor` for details._

32 of the 40 introspectors answer here. Each declares which of three kinds it is: it reads files only, it needs live reflection, or it answers from another source such as the schema dump. What a tier can say is a declaration, so an answer never guesses at what it could not read (ADR-0002). The routes engines and gems draw at boot, and an association a gem adds at boot, are not listed, and the answer says so.

The six rules

The root context file carries six rules the assistant reads before writing code. That file is AGENTS.md, which Cursor, OpenCode and Codex read and CLAUDE.md imports, or CLAUDE.md itself when none of those three is chosen; Copilot reads its own copy in .github/copilot-instructions.md.

1. Verify before you write. Never reference a column, association, route, helper, method, class, partial or gem that a tool call in this turn did not confirm.
2. Mark every assumption with [ASSUMPTION]. "I'd need to check X first" is a good answer.
3. Training data describes average Rails. This app isn't average. When something feels obviously standard, query anyway.
4. Check the inheritance chain before every edit: inherited filters, concerns, includes, STI parents.
5. Empty tool output is information. "0 callers found" means investigate, or ask.
6. Stale context lies. Re-query after writes.

On by default. Turn off with config.anti_hallucination_rules = false if you prefer your own. In an MCP-only or CLI setup, the ground rules prompt gives the assistant the same rules for one session.

The tools

Every tool is read-only and answers from your app. Over MCP each name carries a rails_ prefix (rails_get_schema). From the CLI the short name works: schema, get_schema and rails_get_schema all resolve. About half take detail=summary|standard|full.

Four worth knowing on day one:

• search_code with match_type=trace returns the definition, the source, every caller grouped by file and labelled by kind, and the tests, in one call.
• get_controllers returns the action source with inherited filters, strong params and the render map.
• get_model_details returns associations, validations, scopes, enums and macros from the AST, each tagged [VERIFIED] or [INFERRED].
• query runs read-only SQL inside a read-only transaction, with a timeout, a row limit and column redaction.

Parameters for all 35 are in the tools reference; worked examples in recipes. MCP clients can also read five resource templates, such as rails-ai-context://models/{name} and rails-ai-context://controllers/{name}/{action}, and nine static resources, such as rails://schema and rails://routes.

How it works

[Two columns. What a tool call does: the call arrives over MCP or from the command line, the registry resolves it among 35 auto-registered tools, 40 introspectors read the app with live reflection when it boots and from the routes file, the schema dump and the source when it does not, Prism parses each file once into a cache keyed by content and one walk feeds every listener, each record is tagged VERIFIED, INFERRED or STATIC and the result is cached with a fingerprint of the app, and the answer is composed with schema hints, a cut at the response budget, a tier line and, for most tools, a hint for the next call. What the install writes: the MCP config each chosen tool reads, one root context file, AGENTS.md when Cursor, OpenCode or Codex is chosen and CLAUDE.md otherwise, with a copy for Copilot, rules that load where they apply, six anti-hallucination rules, its own configuration, and nothing you wrote.]

40 introspectors read the app, one per kind of fact: the schema, the models, the routes, the controllers, the views, the jobs, the gems, the conventions and so on. Each returns a hash and is registered under a key, so a preset is a list of keys: :full runs all 40 and :standard runs 17.

The models and the controllers are read with Prism, Ruby's own parser. Each file is parsed once into a cache keyed by its content, and a single walk over the tree feeds every listener at the same time: associations, validations, scopes, enums, callbacks, macros, strong parameters. A listener tags each record by how it was read: [VERIFIED] for a literal, [INFERRED] for an expression that only runs at runtime. Introspection is cached with a fingerprint of the app's files, so an edit to a model is seen on the next call without a restart.

A tool composes its answer from that context, hydrates it with schema hints for the models it touches, cuts it at the response budget, and ends it with a tier line and, for most tools, a hint for the next call. Three transports reach the same registry: MCP over stdio, MCP over HTTP mounted inside the app, and the command line with no server at all.

Architecture has the modules and the request lifecycle, Introspectors the catalogue of listeners and the rule for what is read from the AST, and the ADRs the decisions with their reasons.

What it costs

It is free and MIT licensed. What you pay is tokens, and only when the assistant asks:

• A tool answer is sized by its detail level and cut at max_tool_response_chars, 200,000 characters by default. Start with summary and drill down.
• The root context file is held to 12,000 bytes in its default compact mode, and a rule file to 8,000, since they share the agent's context with your own instructions.
• The last line of every answer, which asks the assistant to offer you an issue when the gem itself answered wrong, costs about 40 tokens. config.report_issue_line = false leaves it out.

Questions you should ask first

Does it send my code anywhere? No. The tools run on your machine, in your app's process or in a child of it. Two things go online: search_docs fetches Rails guides from GitHub when called with fetch: true, and the standalone binary asks rubygems.org for the latest version at most once a day, in a child process that sends nothing about the app. RAILS_AI_CONTEXT_NO_UPDATE_CHECK=1 turns that off, and it is off in CI.

Can it read my secrets? Files that hold them are never served: .env, config/master.key, credentials.yml.enc, config/database.yml, keys and certificates, on a list you can extend. Logs are redacted before they leave the process, get_config shows credential key names without values, and query never returns an encrypted column. Security has every layer.

Can it change my database? No. Every tool is read-only. query allows SELECT, WITH, SHOW, EXPLAIN and DESCRIBE, runs inside a read-only transaction with a statement timeout, caps the rows, redacts secret columns, and on PostgreSQL plans the statement first and refuses a plan that reaches a secret column by another name.

What does it write into my repository? The MCP config for each AI tool you picked, one root context file with a copy for Copilot, rules files per tool, its own config and a .gitignore line. A re-run keeps your lines and adds what is missing; uninstall takes back only what it wrote. A team that wants no AI files in the repo can skip the install: the tools work with none of it.

Is the HTTP endpoint safe? Mounted inside the app it inherits your routing, auth and middleware. Outside development and test it refuses every request with a 403 unless you set config.allow_http_in_production = true, since it has no authentication of its own.

Which AI tools? Claude Code, Cursor, GitHub Copilot, OpenCode and Codex CLI get context files and MCP config written for them. Any other MCP client, such as Claude Desktop, Windsurf, Zed or a JetBrains IDE (Copilot or AI Assistant), gets an entry from rails-ai-context mcp-entry. A tool with no MCP at all gets the CLI setup, where the context files name the shell commands.

Still wondering about something? The FAQ answers the rest: install, the tools, configuration, context files, performance, security, and how it compares to a hand-written CLAUDE.md or pasting schema.rb into the prompt.

What it will not do

• Write to your app or your database from a tool. Every tool reads. The assistant still decides what to do with the facts, and the six rules ask it to verify before writing; nothing enforces that.
• Say more than it read. Without a boot, 32 of 40 introspectors answer and the rest say [UNAVAILABLE]. query, runtime_info and doctor need a booted app, and query a database.
• Support Rails 9. The gemspec caps railties below 9.0, and the standalone binary warns when it boots one. Compatibility has the CI matrix: 17 Ruby and Rails pairs on Linux, plus macOS and Windows legs.
• Time-limit query on SQLite under Windows. There is no fork there, so that query runs without its limit, and the MCP configs start the server through cmd /d /c, since a gem's executables on Windows are batch files.

Install options and configuration

Three setups, picked at install time with a flag, or afterwards with the change the setup prompt:

MCP over stdio is the default. Each AI tool gets its own config file (.mcp.json, .cursor/mcp.json, .vscode/mcp.json and .github/mcp.json, opencode.json, .codex/config.toml) that starts bundle exec rails-ai-context serve, or rails-ai-context serve standalone. MCP over HTTP mounts the server inside your app:

# config/routes.rbmount RailsAiContext::Engine, at: "/mcp" if defined?(RailsAiContext::Engine)

The if keeps the routes file loading where the gem is not installed, as in production, where Bundler skips the development group. There is also a standalone HTTP process: rails-ai-context serve --transport http --port 6029. The CLI runs the same 35 tools with no server:

bin/rails 'ai:tool[search_code]' pattern="publishable?" match_type=tracerails-ai-context tool schema --table users --detail full

The commands, in the Gemfile and standalone:

The binary finds the app the way bin/rails does, so it runs from any subdirectory. Flags shared by the app-reading commands: --app-path PATH to target another directory, --environment ENV to set RAILS_ENV, and --no-boot to answer from source. The full list is in the CLI reference.

A Gemfile install writes an initializer and .rails-ai-context.yml; a standalone install writes the YAML file alone, with the same keys. Every option with its default is in Configuration:

# config/initializers/rails_ai_context.rbif defined?(RailsAiContext) && RailsAiContext.respond_to?(:configure)  RailsAiContext.configure do |config|    config.ai_tools  = %i[claude cursor]   # which AI tools to generate for    config.tool_mode = :mcp                # :mcp (default) or :cli    config.preset    = :full               # :full (40 introspectors) or :standard (17)  endend

# app/mcp_tools/rails_get_business_metrics.rbclass RailsGetBusinessMetrics < MCP::Tool  tool_name "rails_get_business_metrics"  description "Key business metrics for this app"  input_schema(properties: { period: { type: "string" } })
  def self.call(period: "week", server_context: nil)    MCP::Tool::Response.new([{ type: "text", text: "Users this #{period}: #{User.recent.count}" }])  endend
# config/initializers/rails_ai_context.rbconfig.custom_tools = ["RailsGetBusinessMetrics"]

include RailsAiContext::TestHelper
response = execute_tool("business_metrics", period: "month")assert_tool_response_includes(response, "Users")

ActiveSupport::Notifications.subscribe("rails_ai_context.tools.call") do |event|  ms = (event.payload[:duration].to_f * 1000).round  Rails.logger.info "[MCP] #{event.payload[:tool_name]} #{ms}ms"end

To take it out again:

bundle exec rails-ai-context uninstall && bundle remove rails-ai-context   # in the Gemfilerails-ai-context uninstall                                                 # standalone

It removes the context files, the MCP entries, its config, its .gitignore lines and its pre-commit hook, and keeps what you wrote: your lines in CLAUDE.md, other servers in an MCP config, a hook you edited. Or paste the uninstall prompt.

Troubleshooting

bundle exec rails-ai-context doctor (or bin/rails ai:doctor) boots the app, checks every file the install wrote, the MCP configs and the plugins, and prints a readiness score with what to fix. It needs a bootable app, since diagnosing the boot is its job. The fix it prompt hands the same checks to the assistant, and Troubleshooting has the problems people hit most: a server the AI tool never lists, a boot that fails, a Windows path, a bundle that resolves the wrong Rails.

Development

git clone https://github.com/crisnahine/rails-ai-context.gitcd rails-ai-contextbundle installbundle exec rspec              # the suite, on an in-memory SQLite app under spec/internalbundle exec rubocop --parallel # rubocop-rails-omakase stylebundle exec rake e2e           # fresh Rails apps per install path; minutes; TEST_POSTGRES=1 adds a PostgreSQL spec

The website is site/, filled in by scripts/site.rb from the repository: the install prompt from docs/PROMPTS.md, the tool list from the registry, every docs/ page rendered beside it, and the answers recorded by running the gem against site/sample_app. To see it before pushing:

bundle exec ruby scripts/site.rb _sitepython3 -m http.server 8000 -d _site

.github/workflows/pages.yml builds it on every pull request and deploys it from main. CONTRIBUTING.md has the layout of lib/, how to add an introspector, a tool or a Prism listener, and how to run the specs against real database servers.

Contributing and security

Pull requests are welcome; CONTRIBUTING.md has the process, and the code of conduct applies to everyone taking part. Bugs go to the issue tracker with the Ruby, Rails and gem versions and the steps to reproduce. A security problem goes by email, the way SECURITY.md says, and never in a public issue. Only the latest 5.x release gets security fixes.

Learn more

If it saves you a correction loop, star it on GitHub, and if it saves you time, consider sponsoring the project.