@backendkit-labs/mcp-server

MCP (Model Context Protocol) server that exposes code intelligence, skills management, and project context as tools for any MCP-compatible AI client (Claude Desktop, agent-core, Cursor, etc.).

Status: stable — Maintained and production-ready. No new features during the pilot phase.

Table of contents

• Installation
• Running the server
• Tools reference
• Code intelligence index
• Connecting via agent-core
• Connecting via Claude Desktop
• HTTP mode
• Environment variables

Installation

npm install -g @backendkit-labs/mcp-server
# or use via npx:
npx @backendkit-labs/mcp-server

Binary: bk-mcp

Running the server

Stdio mode (default — for Claude Desktop, agent-core)

bk-mcp

# Against a specific project
BK_CWD=/my-project bk-mcp

# With a custom app name
BK_APP_NAME=my-app BK_CWD=/my-project bk-mcp

HTTP mode (multi-session)

bk-mcp --http --port 3100

Tools reference

stack_detect

Detect the project's technology stack using file detection and the code index.

Args:
  cwd  string?  Project directory (defaults to server cwd)

Example:

What stack does this project use?
→ stack_detect()
→ "Detected stack: typescript, node, react, docker"

code_index

Index the codebase into a SQLite graph database. Incremental — only re-indexes changed files.

Args:
  cwd    string?   Project directory
  watch  boolean?  Enable file watcher for live updates (default: false)

Example:

Index this codebase
→ code_index({ watch: true })
→ "Indexed 234 files, 4,821 symbols. File watcher enabled."

What gets indexed: functions, classes, interfaces, types, enums from TypeScript/JS/Go/Python/Java/Kotlin. Callers and callees are recorded for each symbol (call graph).

code_search

Search for symbols by name or keyword.

Args:
  query  string   Symbol name or keyword
  limit  number?  Max results (default 20)

Example:

Where is TokenValidator defined?
→ code_search({ query: "TokenValidator" })
→ "**TokenValidator** (class) — src/auth/token-validator.ts:15
     class TokenValidator { validate(token: string): DecodedToken | null }"

Fuzzy search:

All functions related to payment
→ code_search({ query: "payment", limit: 30 })
→ PaymentProcessor, processPayment, validatePayment, createPaymentIntent, ...

code_context

Given a task description, find the most relevant symbols and return their code, callers, and callees. Bootstraps context for a coding task without manual file reading.

Args:
  task        string   Task or feature description
  max_symbols number?  Max symbols (default 10)

Example:

I need to add rate limiting to the login endpoint
→ code_context({ task: "add rate limiting to login endpoint" })
→ 
### loginHandler (function) — src/routes/auth.ts:67-89
> calledBy: router.post | calls: validateCredentials, createSession
\`\`\`typescript
async function loginHandler(req: Request, res: Response) {
  const { email, password } = req.body;
  ...
}
\`\`\`

The agent immediately knows which files and symbols are relevant.

skills_search

Search available skills by query.

Args:
  query  string   Search term
  cwd    string?  Project directory for stack detection

Example:

Find skills for TypeScript testing
→ skills_search({ query: "typescript testing" })
→ "- **ts-jest-setup** ✓ [global, typescript, ACTIVE]
     Configure Jest for TypeScript with proper transforms
   - **vitest-config** [node, typescript]
     Set up Vitest for unit testing"

✓ means the skill matches the detected stack.

skills_install

Install a skill pack from GitHub or a local path.

Args:
  source  string   GitHub URL or local path
  skill   string?  Install only a specific skill
  pack    string?  Override the derived pack name

Example:

Install our company skills repo
→ skills_install({ source: "https://github.com/my-org/agent-skills" })
→ "Installed pack "backend" with 12 skill(s)."

skills_list

List all installed skill packs.

Args: (none)

Example:

What packs are installed?
→ skills_list()
→ "**global** (builtin)
     Skills: code-review, refactor-pattern, ...
   **my-org-backend** (installed 2026-06-01)
     Skills: stripe-integration, postgres-pattern, ..."

Code intelligence index

The indexer uses Tree-sitter (WASM) to parse source code into a SQLite graph with symbols and their relationships. Storage location:

~/.{BK_APP_NAME}/code-index/{base64url(CWD)}.db

Default: ~/.deepseek-code/code-index/{hash}.db

Supported languages: TypeScript, JavaScript, Go, Python, Java, Kotlin, Rust.

Connecting via agent-core

import { createCodingEngine } from '@backendkit-labs/agent-coding';

const engine = createCodingEngine({
  providers:       { deepseek: { apiKey: process.env.DEEPSEEK_API_KEY! } },
  defaultProvider: 'deepseek',
  mcpServers: [
    {
      name:    'code-intelligence',
      command: 'bk-mcp',
      args:    [],
      env: {
        BK_APP_NAME: 'my-app',
        BK_CWD:      process.cwd(),
      },
    },
  ],
  subAgentContextMessages: 20,
});

// Index then query
await engine.run('Index this codebase');
await engine.run('Find all functions that directly access the database without using the repository layer');

Connecting via Claude Desktop

Add to your Claude Desktop config:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "bk-code": {
      "command": "bk-mcp",
      "env": {
        "BK_APP_NAME": "claude-code",
        "BK_CWD":      "/path/to/your/project"
      }
    }
  }
}

Restart Claude Desktop — the tools will appear in the tool picker.

Multiple projects

{
  "mcpServers": {
    "bk-api": {
      "command": "bk-mcp",
      "env": { "BK_CWD": "/projects/api-server", "BK_APP_NAME": "claude-code" }
    },
    "bk-frontend": {
      "command": "bk-mcp",
      "env": { "BK_CWD": "/projects/frontend", "BK_APP_NAME": "claude-code" }
    }
  }
}

HTTP mode

BK_CWD=/my-project bk-mcp --http --port 3100

Connect:

await mcp.connectStreamableHTTP({
  name:    'code-intelligence',
  baseUrl: 'http://localhost:3100',
});

Environment variables