mcp-broker

mcp-broker

mcp-broker is a local Model Context Protocol process broker for MCP clients.

Think PgBouncer for MCP: one stable local endpoint in front of many upstream MCP servers. The broker owns upstream startup, reuse, cleanup, profile exposure, status, and safe tool routing.

The core idea is simple: do not make every agent session load every upstream tool definition before the user asks a task.

Why this exists

AI coding sessions with many MCP servers tend to accumulate the same problems:

• every client config repeats the same MCP server list
• every new session can start duplicate upstream processes
• OAuth, browser state, local files, and database handles spread across tools
• raw tool lists consume context before the task begins
• hosted connector caches can duplicate local MCP tools
• orphaned MCP processes survive after client sessions exit

mcp-broker puts a small broker facade in front of those upstreams. It is not a hosted workflow builder; it is local infrastructure for keeping MCP clients small, predictable, and under one config contract.

Client profile

| | one local MCP entry v mcp-broker-client | | Unix socket v mcp-broker-daemon | | profile gates, namespace routing, status, cleanup v upstream MCP servers

The client sees a small set of broker tools:

broker_search_tools

broker_describe_tool broker_call_tool broker_status broker_close_session

The upstream MCPs still exist. They are discovered and called through the broker when a task needs them.

Measured context reduction

On 2026-05-24, the measured Codex setup went from many raw MCP and hosted app tool definitions to one broker facade plus a pruned codex_apps cache.

The 92.94% number is a tool-definition count reduction. The 83.65% number is a token reduction for canonical serialized tool payloads measured with tiktoken o200k_base.

See docs/context-reduction-measurement.md for evidence and caveats.

What it does

• Runs one local broker daemon over a Unix socket.
• Exposes one lightweight stdio client shim to MCP clients.
• Starts upstream MCP servers on demand.
• Reuses shared upstreams across sessions when configured.
• Isolates per-session upstreams when state must not be shared.
• Maps upstream tools into stable namespaces.
• Exposes compact search, describe, call, status, and caller-bound session-close tools.
• Enforces profile-level tool budgets and exposure gates.
• Blocks mutating upstream exposure unless a profile allowlist grants it.
• Stores runtime state under $HOME/mcp/mcp-broker, outside the repo.
• Renders MCP client config entries with dry-run, backup, and rollback.
• Provides LaunchAgent install and uninstall flows for macOS.
• Provides Linux systemd user-service render, install, unload, and removal flows.
• Provides Windows PowerShell Scheduled Task render, install, and removal flows.
• Includes unit, journey, live, and e2e tests through Makefile targets.

Core differentiators:

• Profile-scoped exposure: each MCP client gets a configured view of upstreams instead of every tool by default.
• Mutating-tool gates: mutating upstreams stay hidden until a profile allowlist grants access.
• Lifecycle ownership: shared and per-session upstreams are started, watched, stopped, and reaped by the broker.
• Client parity checks: rendered profiles can be validated through the same compact broker facade before config is applied.

Who this is for

Use mcp-broker if you:

• use Codex, Claude Code, AGY CLI, or other MCP clients
• have more MCP tools than you want in every session
• need shared local MCP servers without duplicate process startup
• want one place for OAuth state, browser state, sockets, logs, and cleanup
• need per-client profiles instead of the same tool list everywhere
• want a small broker facade instead of raw upstream tool dumps

This repo is not an enterprise MCP control plane. It is local desktop infrastructure for developer-agent workflows.

Current status

Implemented:

• YAML config loading from config/broker.private.yaml, created from config/broker.example.yaml.
• Strict YAML contract validation for runtime, clients, profiles, upstreams, and policy blocks.
• Public JSON Schema validation through config/broker.schema.json.
• Runtime path derivation from runtime.root.
• Tool namespace mapping from configured upstream prefixes.
• Local upstream subprocess lifecycle management and process-group cleanup.
• Broker daemon over Unix socket.
• MCP client shim and renderers for configured MCP client profiles, including Codex, Claude, and AGY.
• AGY profile rendering to .gemini/config/mcp_config.json, including its MCP allowed-server policy.
• Dry-run client config rendering, apply-time backups, and rollback.
• LaunchAgent render and install scripts with dry-run defaults.
• Compact broker facade for search, describe, call, status, and caller-bound session close.
• Profile validation from YAML smoke probes.
• Discovery parity checks between compact client profiles.
• Public and maintainer quality gates through Makefile targets.
• Shared-runtime guardrail contracts through P3.8 E2E proof while hosted execution remains disabled by default.

Wiring status:

• Codex is wired through the broker.
• Claude is wired through the broker after profile validation and manual /mcp acceptance.
• AGY is wired through the broker by rendering .gemini/config/mcp_config.json.

Public release status:

• The repo is designed to stay public-safe.
• Private upstream inventory, account paths, OAuth state, secrets, sockets, logs, and generated client configs stay outside git.
• Stable release metadata is validated by make release-version-check; publication proof is tracked in docs/distribution.md.
• Docker image support is available for container-friendly configs. Docker MCP Catalog submission still requires Docker review.
• MCPB metadata is present at mcpb/manifest.json for local directory review.

See ROADMAP.md for public-facing release work.

Architecture

mcp-broker has three runtime layers:

The config file is the contract. Profiles decide exposure, upstreams define transport and lifecycle behavior, and smoke probes define safe read calls for validation.

Comparison

Screenshots Or GIF

The quickstart flow should look like this:

make config-init

make config-validate make broker-status make codex-facade-smoke

In an MCP client, /mcp should show one mcp-broker entry. Use broker_status to inspect profile-visible upstream state.

Quickstart

Prerequisites:

• macOS with launchctl for LaunchAgent use.
• Python 3.10 or newer available as python3.
• make.
• Node.js and npx for npm-based upstream MCP servers.
• A clone of this repo.

Package installs:

pipx install mcp-broker

uv tool install mcp-broker brew tap ${HOMEBREW_TAP_REF} brew install mcp-broker

Homebrew installs the same console scripts as the Python package. Package installs do not write MCP client config; client wiring stays an explicit Makefile action.

Docker is for container-friendly configs:

docker build -t mcp-broker:local .

docker run --rm -i mcp-broker:local

Local stdio clients and MCPB-style installs use the package-owned lifecycle:

mcp-broker stdio --init-if-missing

Create the local venv, install dependencies, and verify runtime layout:

make setup

Create private config from the public template:

make config-init

config-init creates the destination directory when needed and copies the public template as the starting point. It does not import local MCP inventory, user paths, or secrets.

Edit config/broker.private.yaml for local upstreams. Keep secret values out of config. Use environment variable names or files under:

$HOME/mcp/mcp-broker/secrets/

Run the quality gate:

make quality-gate

Validate the configured YAML contract:

make config-validate

Start the broker:

make broker-start

Check status:

make broker-status

For the full install flow, see docs/install.md. For a clone-to-running adoption flow, see docs/adoption-guide.md#clone-to-running-path. For shared-runtime boundaries, see docs/shared-runtime-guardrails.md. P3.8 E2E proof covers tenant isolation, authz denial, quota denial, session affinity, audit events, rollback, degraded mode, local-only routing, and shared-eligible routing. Hosted execution remains unsupported in the public local broker.

Runtime layout

Default runtime root:

$HOME/mcp/mcp-broker/

|- backups/ |- logs/ |- renders/ |- run/ |- secrets/ |- sockets/ - state/ - upstreams/

Runtime files are not repo files. Upstream OAuth state, browser state, secret files, sockets, logs, rendered client configs, backups, and daemon state belong under the runtime root.

See docs/runtime-layout.md.

Client wiring

Back up a client config:

make config-backup CLIENT=codex

Dry-run render:

make config-render CLIENT=codex CONFIG_RENDER_APPLY=0

Apply after reviewing the rendered file under $HOME/mcp/mcp-broker/renders/:

make config-render CLIENT=codex CONFIG_RENDER_APPLY=1

Rollback:

make config-rollback CLIENT=codex

Use CLIENT=claude or CLIENT=agy after that profile smoke passes and that client is intended to use the broker. For new JSON-based MCP clients, generate a starter block:

make profile-snippet NEW_PROFILE=local-client NEW_CLIENT_FORMAT=mcp-settings-json

See docs/add-profile.md for the full new-profile flow.

Compact broker facade

The compact facade keeps chat-facing profiles small:

| broker_call_tool | Call one upstream tool through br