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:

Codex /mcp shows the single mcp-broker entry by design. Per-upstream visibility, status, and socket path come from broker_status.

Profiles and safety

Profiles decide which upstreams a client can see and call.

Supported concepts:

• max_tools protects clients from huge tool lists.
• compact_tools_enabled exposes broker facade tools instead of raw upstream tools.
• broker_tool_name_style adapts broker facade names for clients that cannot surface dotted tool names.
• mcp_allowed_servers renders client settings for MCP clients that require an explicit server allowlist.
• allow_mutating_upstreams is required before a mutating upstream can be exposed.
• shared mode reuses one upstream process where shared account state is acceptable.
• per_session mode isolates upstream state per client session.
• per_call mode starts a fresh stdio process for each tool call or tool-list operation, then stops that process before returning.
• disabled mode keeps compatibility records without exposing the upstream.

Status reports active_call_count for in-flight per_call operations, including tool discovery. It returns to zero after success, failure, or timeout cleanup. This count is separate from session_count and is not a count of saved threads. Reading status does not start an upstream; profile-hidden upstreams stay hidden.

Protected surfaces such as OAuth, browser state, filesystem roots, and databases require explicit config and validation. Public examples stay disabled or placeholder-based.

See docs/security-review.md and docs/upstream-compatibility-matrix.md. For a deeper safety checklist, see docs/safety.md.

Config contract

The public template is config/broker.example.yaml. The matching JSON Schema is config/broker.schema.json.

Supported top-level sections:

schema_version: 1
runtime: {}
broker: {}
profiles: {}
clients: {}
upstreams: {}

The loader rejects unknown keys. Runtime placeholders such as {runtime.root}, {runtime.state_dir}, and {runtime.secrets_dir} can be used in upstream command, args, working directory, and env file paths.

make config-validate checks the selected CONFIG_PATH against the public JSON Schema first, then runs the runtime loader so semantic rules are enforced from the same code path the broker uses.

Each enabled upstream exposed to a profile should define a safe smoke probe:

smoke:
  query: read example graph
  tool: example-store.read_graph
  arguments: {}
  call: true

make profile-validation PROFILE=<profile> validates every enabled upstream visible to that profile through broker_status, broker_search_tools, broker_describe_tool, and the configured safe broker_call_tool.

Codex operator acceptance

Repo-owned tests validate broker behavior through the local client shim. The last Codex-specific check has to run inside an active Codex session because that is where the deferred MCP wrapper tools exist.

Generate the current acceptance steps from YAML:

make codex-deferred-acceptance

The target reads the configured smoke probes and prints the exact mcp__mcp_broker__ wrapper calls for search, describe, and safe call. It does not invoke Codex, does not call an external LLM session, and is not part of make quality-gate.

See docs/codex-deferred-tool-acceptance.md.

LaunchAgent

Render without writing:

make launchagent-install

Apply and load:

make launchagent-install LAUNCHAGENT_APPLY=1
make launchagent-load
make broker-status

Unload or remove:

make launchagent-unload
make launchagent-uninstall LAUNCHAGENT_APPLY=1

systemd

Linux user-service install uses the same runtime root and config path contract:

make systemd-install
make systemd-install SYSTEMD_APPLY=1
make systemd-load

For package installs, set MCP_BROKER_DAEMON_COMMAND to the installed daemon path before applying the service.

Windows

Windows startup uses PowerShell Scheduled Task commands with the same runtime root and config path contract:

make windows-install
make windows-install WINDOWS_APPLY=1
make windows-load

Remove it with:

make windows-unload
make windows-uninstall WINDOWS_APPLY=1

Test and release gates

Run all test tiers:

make test

Run the public quality gate:

make quality-gate

The coverage gate uses line and branch coverage for Python source.

Run the release gate when preparing a tag:

make release-gate

release-gate runs package, smoke, and mutation checks. Mutation receives a release-scoped child count derived from LOCAL_CPU_BUDGET and RELEASE_GATE_JOBS, so it does not take the full CPU budget while other release children run. Mutation runs public unit and journey tests last and writes var/quality/mutation_stats.json with total counts, score, and ranked blocked_by_file entries. On macOS, the release gate runs mutation inside a Linux container to avoid local mutmut fork failures. E2E tests remain in make quality-gate.

Run smoke and runtime cleanup checks:

make config-validate
make broker-smoke
make broker-stop
make broker-reap
make doctor
make release-smoke

Release or client config apply should wait for:

• make quality-gate
• make config-validate
• make broker-smoke
• dry-run config render for each intended client
• rollback test
• make release-smoke
• make release-gate before tagging
• make doctor with no stale broker-owned resources

See docs/release-checklist.md.

Public commands

These targets use this repo plus declared Python and Node prerequisites:

For source contributions, install gitleaks on PATH and run make hooks-install in each checkout. The tracked pre-commit hook requires a redacted staged secret scan, then runs commit-tier affected tests through Make. Missing scanner tools, empty staged scope, and scan or test failures block the commit. GITLEAKS and PYTHON select the scanner and maintainer interpreter. Installation uses Git's worktree-specific config and refuses unknown existing hooks or common core.worktree/bare settings that need migration. No hosted workflow is started.

make setup
make config-init
make test
make test-unit
make test-journey
make test-live
make test-e2e
make test-cov
make precommit
make quality-gate
make release-gate
make config-validate
make broker-smoke
make broker-start
make broker-status
make broker-stop
make broker-reap
make doctor
make config-backup
make codex-app-policy
make config-render
make config-rollback
make tools-count
make facade-smoke
make codex-facade-smoke
make claude-facade-smoke
make agy-facade-smoke
make profile-validation
make codex-profile-validation
make claude-profile-validation
make agy-profile-validation
make discovery-parity
make codex-claude-discovery-parity
make codex-deferred-acceptance
make launchagent-install
make launchagent-load
make launchagent-unload
make launchagent-uninstall
make systemd-install
make systemd-load
make systemd-unload
make systemd-uninstall
make windows-install
make windows-load
make windows-unload
make windows-uninstall
make linux-container-smoke
make windows-powershell-smoke
make release-smoke
make mutation
make mutation-linux

make quality-gate is repo-local. It does not call personal scripts outside this repo.

make codex-deferred-acceptance is maintainer-only. It does not invoke Codex or an external LLM session. It reads the same YAML smoke probes and prints the exact mcp__mcp_broker__ deferred wrapper calls to run inside an active Codex session. See docs/codex-deferred-tool-acceptance.md.

Project tree

mcp-broker/
|- .gitignore
|- Makefile
|- README.md
|- pyproject.toml
|- requirements.txt
|- config/
|  |- broker.example.yaml
|  |- broker.private.yaml        # local, ignored by git
|  `- broker.schema.json
|- docs/
|- registry/
|- scripts/
|  `- check_mutation_stats.py
|- src/
|  `- mcp_broker/
|- tests/
|  |- unit/
|  |- journey/
|  |- live/
|  |- e2e/
|  `- support/
`- var/                         # tracked skeleton; generated contents ignored

Generated reports stay under var/, especially var/coverage/, var/test-logs/, and var/quality/.

Docs

• SECURITY.md
• CONTRIBUTING.md
• ROADMAP.md
• docs/install.md
• docs/add-profile.md
• docs/migration.md
• docs/adoption-guide.md
• docs/comparison.md
• docs/distribution.md
• docs/github-publication.md
• docs/community-launch.md
• docs/auth-recipes.md
• docs/architecture.md
• docs/protocol.md
• docs/runtime-layout.md
• docs/safety.md
• docs/security-review.md
• docs/upstream-compatibility-matrix.md
• docs/codex-deferred-tool-acceptance.md
• docs/context-reduction-measurement.md
• docs/mutation-testing.md
• docs/release-checklist.md
• docs/troubleshooting.md

Design rules

• Keep upstream definitions in central config.
• Keep runtime state under $HOME/mcp/mcp-broker.
• Keep private upstream inventory in config/broker.private.yaml, which is ignored by git.
• Keep secret values out of config and source.
• Do not hardcode personal paths in source, tests, docs, or public config.
• Run build, test, runtime, and config operations through the Makefile.