Claude Mcp Bridge — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited Claude Mcp Bridge (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
MCP server that lets any agent or MCP host (Cursor, Codex, Antigravity, another Claude Code, your own app) escalate to Opus via Claude Code running headless (claude -p / claude --bg).
Opus is expensive but very capable — this bridge is deliberately narrow: complex reasoning, architecture questions, adversarial review, and autonomous work. Cheap mechanical work (mapping a repo, scanning files) should stay on the host agent's lighter model.
Nested MCP delegation is blocked: Opus sessions spawned by this bridge cannot call claude-mcp-bridge (or any MCP) again.
| Tool | Purpose | Model |
|---|---|---|
ask | Read-only Q&A — architecture, explanation, planning (claudecode({ prompt: "..." })) | Opus (forced) |
delegate | Full autonomous task — edit, shell, tests (max turns, synchronous) | Opus (forced) |
fire | Fire-and-forget background task — returns session_id immediately | Opus (forced) |
check | Compact progress report for a fire session | read-only (no model) |
follow_up | Continue a prior session by session_id | Opus (forced) |
fire({ prompt: "Refactor the auth module to use JWT" })
→ returns session_id immediately
check({ session_id: "..." })
→ check progress anytime// Ask Opus to explain architecture (read-only)
{ "prompt": "Explain the architecture of this project" }
// Delegate autonomous work (waits for completion)
{ "prompt": "Refactor auth to JWT and fix all tests" }
// Background refactor — poll while you keep working
{ "prompt": "Refactor the auth module to use JWT", "name": "JWT refactor" }
// → session_id: ...
{ "session_id": "..." }Every tool accepts: cwd (project root), effort (low \| medium \| high, default high).
claude CLI v2.1.139+ installed and authenticated (claude once to log in).Background tasks (fire / check) require claude --bg and claude agents --json.
git clone https://github.com/JaimeJunr/claude-mcp-bridge.git
cd claude-mcp-bridge
npm install
npm run buildClaude Code:
claude mcp add claude-bridge -s user -- node /abs/path/to/claude-mcp-bridge/dist/index.jsCursor / Codex / any host — add to its mcp.json:
{
"mcpServers": {
"claude-bridge": {
"command": "node",
"args": ["/abs/path/to/claude-mcp-bridge/dist/index.js"]
}
}
}| Var | Default | Meaning | |||
|---|---|---|---|---|---|
CLAUDE_BIN | claude | Path to the claude binary. | |||
CLAUDE_BRIDGE_MODEL | opus | Model alias — always Opus in this bridge. | |||
CLAUDE_BRIDGE_EFFORT | high | Default reasoning effort. | |||
CLAUDE_BRIDGE_MAX_TURNS | 20 | Max agentic turns for delegate / fire. | |||
CLAUDE_BRIDGE_PERMISSION | acceptEdits | Permission mode for delegate/fire: plan \ | default \ | acceptEdits \ | bypassPermissions. |
CLAUDE_BRIDGE_TIMEOUT_MS | 600000 | Per-call timeout for sync tools (ask, delegate). | |||
CLAUDE_BRIDGE_SHORT_TIMEOUT_MS | 30000 | Timeout for fire / check. |
Security:acceptEditslets delegated Opus edit files autonomously;bypassPermissionsadditionally allows unrestricted shell. Only raise this for trusted workspaces.askalways usesplan(read-only).
npm test # vitest — unit tests for arg building / json parsing
npm run dev # run from source via tsxMIT
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.