@rudderjs/boost

AI developer-experience layer for Rudder. Exposes your project's internals to AI coding assistants via MCP — Claude Code, Cursor, GitHub Copilot, Codex CLI, Gemini CLI, and Windsurf. Also ships per-package AI coding guidelines and task-specific skill modules that are concatenated into your agent's instructions on install.

Installation

New projects can opt in via create-rudder — Boost appears in the optional-package multiselect during scaffolding. To add it to an existing project:

pnpm add -D @rudderjs/boost

// bootstrap/providers.ts
import { BoostProvider } from '@rudderjs/boost'

export default [
  // ...other providers
  BoostProvider,
]

Quick start

pnpm rudder boost:install                        # auto-detect agents, generate configs
pnpm rudder boost:install --agent=claude-code    # specific agent
pnpm rudder boost:install --agent=cursor,copilot # multiple agents

boost:install detects which agents you use by checking for existing config files. Defaults to Claude Code if nothing is detected.

Supported agents

Agent	Guidelines file	MCP config	Skills
Claude Code	`CLAUDE.md`	`.mcp.json`	`.ai/skills/`
Cursor	`.cursorrules`	`.cursor/mcp.json`	`.ai/skills/`
GitHub Copilot	`.github/copilot-instructions.md`	`.vscode/mcp.json`	—
Codex CLI	`AGENTS.md`	`.mcp.json`	—
Gemini CLI	`GEMINI.md`	`.gemini/settings.json`	—
Windsurf	`.windsurfrules`	`.windsurf/mcp.json`	—

Commands

pnpm rudder boost:install                          # generate per-agent configs, guidelines, skills, boost.json
pnpm rudder boost:install --agent=cursor,copilot   # restrict to specific agents
pnpm rudder boost:install --include-all-skills     # install every shipped skill, even ones whose target package isn't present
pnpm rudder boost:update                           # re-scan packages and update all agent configs
pnpm rudder boost:mcp                              # start the MCP server (stdio transport)

boost:install writes a boost.json at the project root recording the agents you selected and the skills it installed; boost:update reads it back to drive incremental re-scans without re-prompting. Commit boost.json so teammates and CI get the same agent configuration.

Skill targeting (appliesTo)

A skill's frontmatter can declare appliesTo: [<package>, ...]. boost:install only installs the skill when at least one of those packages is present in the project — keeps the skills directory focused on what the project actually uses. Use --include-all-skills to bypass and install every shipped skill, even ones whose target package isn't installed.

# boost/skills/orm-models/SKILL.md
---
name: orm-models
appliesTo:
  - @rudderjs/orm
  - @rudderjs/orm-prisma
trigger: when defining or editing ORM models
---

MCP tools

Boost ships an MCP server that runs alongside your app. AI agents connect and call these tools to get live context:

Tool	Description
`app_info`	Installed `@rudderjs/*` packages, versions, Node.js version, package manager
`db_schema`	Native typed registry (`.rudder/types/models.d.ts`) or Prisma schema — parsed models with fields/types, or raw source
`route_list`	All HTTP routes with methods, paths, middleware, source files
`model_list`	ORM models in `app/Models/` with table names and field types
`config_get`	Read config files — list all or read a specific one by key
`last_error`	Latest log entries from `storage/logs/`
`db_query`	Execute read-only SQL SELECT queries via Prisma
`read_logs`	Read log entries with filtering by level and search term
`browser_logs`	Read browser console logs from Vite dev server
`get_absolute_url`	Convert relative URI paths to absolute URLs using APP_URL
`search_docs`	Search local `@rudderjs/*` package documentation by keyword
`commands_list`	List all rudder commands (built-in + package + user-defined) with args, options, and source. Optional `namespace` filter
`command_run`	Execute a rudder command and return stdout, stderr, exit code, and duration

MCP resources

Guidelines are exposed as MCP resources so agents can re-read them on demand:

Resource URI	Description
`guidelines://orm`	Guidelines for `@rudderjs/orm`
`guidelines://auth`	Guidelines for `@rudderjs/auth`
`guidelines://all`	All package guidelines concatenated
...	One resource per installed package with `boost/guidelines.md`

AI guidelines & skills

Each @rudderjs/* package can ship boost/guidelines.md and boost/skills/*/SKILL.md files. Running boost:install collects them into your project:

.ai/guidelines/{package}.md — per-package AI coding guidelines
.ai/skills/*/SKILL.md — on-demand task-specific knowledge modules
Per-agent guideline files (CLAUDE.md, .cursorrules, etc.) — concatenated guidelines

Your app can ship its own boost/guidelines.md at the project root to include app-specific context (naming conventions, architecture decisions, business rules) — it gets merged in alongside the package guidelines.

Custom agent registration

Third-party packages or users can register custom agent adapters:

import { Boost } from '@rudderjs/boost'
import type { BoostAgent } from '@rudderjs/boost'

const myAgent: BoostAgent = {
  name:               'my-ide',
  displayName:        'My IDE',
  detect:             (cwd) => existsSync(join(cwd, '.my-ide')),
  supportsGuidelines: true,
  supportsMcp:        true,
  supportsSkills:     false,
  installGuidelines:  async (cwd, content) => { /* ... */ },
  installMcp:         async (cwd, cmd) => { /* ... */ },
}

Boost.registerAgent(myAgent)

Custom agents appear in boost:install auto-detection and --agent= selection.

Programmatic use

import { getAppInfo, getDbSchema, getRouteList, getModelList, searchDocs } from '@rudderjs/boost'

const info    = getAppInfo(process.cwd())
const schema  = getDbSchema(process.cwd())
const routes  = getRouteList(process.cwd())
const models  = getModelList(process.cwd())
const results = searchDocs(process.cwd(), 'middleware')

Useful for custom dashboards, CI checks, or embedding project context into your own tooling.

Notes

boost:mcp requires a fully bootstrapped application — it reads live routes, models, and config from the running DI container.
db_schema reads the native typed registry (.rudder/types/models.d.ts, written by rudder migrate) or the Prisma schema (single- or multi-file).
last_error returns the most recent entries from storage/logs/ filtered by level.
Works with any MCP-compatible AI client beyond the listed agents — the standard MCP protocol is the wire format.