Skip to main content
Glama

AI Bridge

Local-first MCP hub that turns the AI coding CLIs you already have into one dispatchable tool fleet.

CI License: MIT Node MCP deps

AI Bridge is a resident HTTP service (default 127.0.0.1:9800) that sits between your agent and the AI tools already installed on your machine. It registers external coding CLIs — Claude Code, OpenCode, Qoder, Atom Code, OpenClaw, WorkBuddy, Codely, Agnes MCP and more — as dispatchable tools, exposes them through a 16-tool MCP server, and serves OpenAI- and Anthropic-compatible endpoints so existing clients can reach the same fleet unchanged.

Runtime footprint: one dependency (ws), ~15,300 lines of JavaScript across 42 files, no build step.


Why

Problem

What AI Bridge does

Many heterogeneous AI CLIs, each with its own invocation shape

One registry + one dispatch verb, with auto-fallback across tools

Agents can't see what ran, what it cost, who called it

Traffic capture, token accounting, per-client ACL, structured logs

Parallel agent sessions trample each other's files

Per-task workspace isolation with create/list/restore

Cloud relays are a privacy and latency liability

Loopback-first: binds 127.0.0.1 by default, no cloud dependency

Clients speak different protocols

MCP (stdio bridge) + OpenAI /v1/chat/completions + Anthropic /v1/messages

Related MCP server: Commons

Quick start

npm install          # installs the single dependency (ws)
node server.js       # listens on 127.0.0.1:9800

Point an MCP client at the stdio bridge:

// claude_desktop_config.json / .cursor/mcp.json
{ "mcpServers": { "ai-bridge": { "command": "node", "args": ["<repo>/bridge-stdio.js"] } } }

Health check:

curl http://127.0.0.1:9800/api/health
# {"status":"ok","uptime":…,"port":9800,"clients":…,"version":"1.1"}

Optional Windows service install: install-service.ps1 / install-service.bat (node-windows or NSSM); register-autostart.ps1 for logon start. Uninstall: uninstall-service-nodewin.js.

MCP tools (16)

bridge-stdio.js exposes the hub to any MCP client:

Tool

Purpose

dispatch

Send a task to one or more registered CLI tools (auto-fallback)

task_status

Poll an async task until a terminal state

list_tools / toggle_tool / list_switches

Inspect and switch the tool fleet on/off

list_clients / manage_client

Registry and ACL for calling agents (register/approve/deny/kick)

handoff

Transfer an in-flight task between tools

discuss

Multi-model discussion session (inject / interrupt / auto-stop)

create_workspace / list_workspaces

Isolated per-task working directories

logs

Structured event log query

memory_search / memory_recent

Cross-session memory retrieval

token_stats

Token accounting per tool/day

recognize_image

Vision pass-through to a local OCR/vision endpoint

HTTP surface

47 /api/* routes plus protocol-compatible endpoints. Grouped:

Group

Routes

Health & metrics

/api/health, /api/metrics, /api/stats, /api/db/stats, /api/insights

Tool fleet

/api/tools, /api/toggle, /api/discover-models, /api/models/sync, /api/refresh

Dispatch & tasks

/api/dispatch, /api/task/<id>, /api/result, /api/poll, /api/handoff

Discussions

/api/discuss, /api/discuss/status/<id>, /api/discuss/active, /api/discuss/inject, /api/discuss/interrupt, /api/meetings

Memory

/api/memory/search, /api/memory/recent, /api/memory/experience, /api/memory/stats

Observation

/api/observe/status, /api/observe/route, /api/observe/takeover, /api/observe/restore, /api/mitm/status

Traffic & usage

/api/traffic, /api/traffic/capture, /api/traffic/clear, /api/usage/report, /api/telemetry

Clients & ACL

/api/client/register, /api/client/unregister, /api/clients, /api/clients/approve, /api/clients/deny, /api/clients/kick

Workspaces

/api/workspaces, /api/workspaces/<id>

Compatibility

/v1/chat/completions, /v1/models, /v1/messages, /proxy/<source>/<key>/…

Security model

Defaults are chosen for a single-user workstation; every claim below is backed by code:

  • Loopback by default. AI_BRIDGE_HOST defaults to 127.0.0.1. The peer check reads req.socket.remoteAddress at exactly one place — isLoopbackPeer() (server.js:1112) — so /api/*, /v1/* and /proxy/* cannot disagree about who is local.

  • Constant-time credential compare. checkAuth() (server.js:1123) compares length first, then crypto.timingSafeEqual, avoiding the RangeError trap on unequal lengths.

  • Unauthenticated relaying is refused off-loopback. The proxy gate is isLoopbackPeer(req) || (!!CONFIG.AUTH_TOKEN && checkAuth(req)) (server.js:1195): with no token configured, non-loopback callers cannot turn the hub into an open LLM relay. Spoofed X-Forwarded-For: 127.0.0.1 is ignored — only the socket peer counts.

  • Per-client ACL. Callers register and must be approved; deny / kick revoke. Pending registrations are tracked separately from the active registry.

  • Runtime data never enters version control. .gitignore excludes logs.jsonl*, traffic.jsonl, token-stats.json, memory.json, switches.json, clients-acl.json, workspaces.json, .env, *.local, and the locally-retained codely-client.js.

  • See SECURITY.md for the threat model and reporting channel.

Configuration

32 environment variables, all optional. Grouped:

Group

Variables

Network

AI_BRIDGE_HOST, AI_BRIDGE_PORT, AI_BRIDGE_TOKEN, AI_BRIDGE_CORS, AI_BRIDGE_BASE_URL

Storage

AI_BRIDGE_DATA_DIR, AI_BRIDGE_DB, AI_BRIDGE_OBSERVE_DIR, AI_BRIDGE_WORKSPACE_ROOT, AI_BRIDGE_TOOLS_PATH

Dispatch

AI_BRIDGE_MAX_CONCURRENT_DISPATCH, AI_BRIDGE_EXECUTE_TIMEOUT, AI_BRIDGE_HTTP_TIMEOUT, AI_BRIDGE_PROXY_TIMEOUT, AI_BRIDGE_CAP_EXEC, AI_BRIDGE_TOOLSET, AI_BRIDGE_RESULT_MAX_CHARS

Clients

AI_BRIDGE_CLIENT, AI_BRIDGE_CLIENT_TTL

Discussions

AI_BRIDGE_MEETING_MAX_MS, AI_BRIDGE_MEETING_MAX_TOKENS

Sampling & cache

AI_BRIDGE_SAMPLING_MAX_TOKENS, AI_BRIDGE_SAMPLING_TIMEOUT, AI_BRIDGE_CACHE_TTL, AI_BRIDGE_CACHE_MAX, AI_BRIDGE_DAILY_TOKEN_BUDGET

Upstream & vision

AI_BRIDGE_LOCAL_URL, AI_BRIDGE_CLOUD_URL, AI_BRIDGE_OLLAMA_BASE, AI_BRIDGE_VISION_HOST, AI_BRIDGE_VISION_PORT, AI_BRIDGE_VISION_MODEL

Testing

Eight .mjs suites plus legacy .js suites:

npm test        # test-unit + test-workspace + test-store-corrupt + test-e2e
npm run lint    # eslint over .js and .mjs

Suite

Covers

test-unit.mjs

Core units, tool registry, routing score

test-workspace.mjs

Workspace isolation, create/list/restore

test-store-corrupt.mjs

Persistence under corrupted/partial state

test-e2e.mjs

End-to-end dispatch against a hub instance

test-external-fixes.mjs

Off-loopback auth and proxy gating

test-audit-fixes.mjs

Regression locks for audited defects

test-progress-watch.mjs

Async task progress states

test-ledger-hygiene.mjs

Internal bookkeeping freshness (optional)

⚠️ test-e2e.mjs binds a fixed port. Do not run the suite while a production hub is listening on it.

CI runs lint and the test suites on push and pull request: .github/workflows/ci.yml.

Extending

Tool adapters live in adapters/:

File

Role

base.js

Adapter contract (detect / execute / normalize)

cli-adapter.js

Spawn a local CLI, capture stdout, normalize the result

http-adapter.js

Call an HTTP tool endpoint

Adding a tool = one adapter module + a tools.json entry (id, name, command/commandPaths, enabled, available).

Limitations and non-goals

  • Windows-first packaging. Service install scripts target node-windows/NSSM; on Linux/macOS run node server.js under your own supervisor.

  • Compatibility shim, not a full API. /v1/* covers chat completions, model listing and Anthropic messages — enough for clients and tests, not every upstream feature.

  • Ships no models. AI Bridge orchestrates tools you already have; it does not host weights.

  • Single-user by design. The ACL separates callers, but there is no multi-tenant quota or billing layer.

  • QiDi-Agent — the orchestration framework (task splitting, interface contracts, three-layer quality checks, multi-path merge) this hub was extracted from. AI Bridge is the runtime; QiDi-Agent is the framework layer.

License

MIT © Qidi AI — see LICENSE.


AI Bridge 是一个常驻本机 HTTP 服务(默认 127.0.0.1:9800)的 MCP 中枢:把你机器上已装的各种 AI 编程 CLI(Claude Code、OpenCode、Qoder、Atom Code、OpenClaw、WorkBuddy、Codely、Agnes MCP 等)注册成可派发的工具,通过 16 个 MCP 工具暴露给任意 MCP 客户端,同时提供 OpenAI / Anthropic 兼容端点,让现有客户端不改代码就能调用同一批工具。

特点:依赖只有一个 ws、约 15,300 行 JS、无需构建;默认只绑回环、凭据恒定时间比较、非回环无凭据拒绝透传;带流量留痕、token 记账、客户端 ACL、按任务隔离工作区、多模型讨论会话(可注入/中断/自动止损)。

它是 QiDi-Agent 框架拆出来的运行时层:QiDi-Agent 负责任务拆分、接口契约、三层质检与多路合并,AI Bridge 负责把这些能力变成可被任意 Agent 调用的服务。

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables orchestrating multiple AI coding tools (Claude Code, Codex CLI, Gemini CLI) on shared repositories with file locking, knowledge capture, task planning, and drift detection via an MCP server.
      166
      -
    • A
      license
      B
      quality
      B
      maintenance
      Lets multiple AI agents and parallel coding sessions share one memory and task ledger, so they can claim work instead of duplicating it, resume each other's handoffs and reasoning, and use credentials without ever seeing the keys. Connects assistants such as Claude Code, Codex, Cursor, Gemini, ChatGPT and custom agents over MCP, backed by a single local SQLite file or a self-hosted Postgres hub.
      41
      1
      MIT
    • F
      license
      Not graded
      quality
      A
      maintenance
      This local MCP server lets a primary AI assistant delegate tasks in parallel to specialized sub-agent CLIs (Gemini, Claude, Codex, Cursor) that share the same project context, then aggregates and synthesizes their findings into one response. It uses existing CLI tools and API subscriptions with no markup, supporting 26+ MCP-compatible IDEs.
      -
    • F
      license
      Not graded
      quality
      A
      maintenance
      Enables a primary AI coding assistant to delegate work in parallel to multiple specialized AI CLI tools (Gemini, Claude, Codex, Cursor) sharing the same project context, then aggregates their findings into a single synthesized response. Runs locally as an MCP server with zero markup, using existing CLI tools, API keys, and subscriptions across 26+ IDE clients.
      -