Mcp Gateway

MCP Gateway by unrelated.ai

MCP infrastructure for turning existing systems into MCP servers and serving them safely at scale.

[](https://github.com/unrelated-ai/mcp-gateway/actions/workflows/ci.yml) [](https://github.com/unrelated-ai/mcp-gateway/actions/workflows/security-rustsec.yml) [](https://github.com/unrelated-ai/mcp-gateway/actions/workflows/security-cargo-deny.yml) [](https://github.com/unrelated-ai/mcp-gateway/actions/workflows/security-trivy-adapter.yml) [](https://github.com/unrelated-ai/mcp-gateway/actions/workflows/security-trivy-gateway.yml) [](https://github.com/unrelated-ai/mcp-gateway/actions/workflows/security-trivy-operator.yml) [](https://github.com/unrelated-ai/mcp-gateway/actions/workflows/security-trivy-migrator.yml) [](https://github.com/unrelated-ai/mcp-gateway/actions/workflows/security-trivy-ui.yml) [](LICENSE)

Quick Start (recommended): Docker (published images) + Web UI

If you just want to try the product experience, this is the fastest path (no repo clone needed).

This example pins component versions (adjust as desired):

• Gateway + migrator: 0.12.4
• Web UI: 0.8.3

1. Download a Compose file that uses published images:

curl -fsSL -o mcp-gateway-compose.yml \
  https://raw.githubusercontent.com/unrelated-ai/mcp-gateway/main/docker-compose.quickstart.yml

1. Start the stack:

GATEWAY_VERSION=0.12.4 UI_VERSION=0.8.3 \
  docker compose -f mcp-gateway-compose.yml up -d

1. Open the Web UI:

• http://127.0.0.1:27102

On a fresh install (bootstrap enabled + empty DB), onboarding will guide you through creating the first tenant and profile.

1. Connect your MCP client to the profile endpoint you created in the UI:

• http://127.0.0.1:27100//mcp

Useful commands:

• docker compose -f mcp-gateway-compose.yml down (stop the stack)
• docker compose -f mcp-gateway-compose.yml down -v (also wipe the DB volume)

Quick Start (from source): Docker Compose + Web UI

If you're developing this repo locally, the Makefile-driven workflow is convenient:

make up

Useful commands:

• make down (stop the stack)
• make up-reset (wipe the demo DB; deletes all tenants/config; then run make up again to re-trigger onboarding)

Tip: this repo uses a Makefile heavily—run make help to see all targets.

Why this exists

MCP clients need to call tools, but most real systems already exist as:

• HTTP APIs
• existing stdio MCP servers (node/python/etc)
• multiple internal services that need to be aggregated into one tool surface

This workspace makes those systems consumable via streamable HTTP MCP without rewriting them into a bespoke MCP server.

What this ships

• Adapter (unrelated-mcp-adapter): expose HTTP/OpenAPI/stdio MCP as one MCP server at /mcp
• Gateway (unrelated-mcp-gateway): tenant/profile-based MCP proxy at /{profile_id}/mcp with auth and policy
• Admin tooling:
• CLI (unrelated-gateway-admin)
• Tenant-level Web UI (beta)

Key concepts

| Concept | What it is | Why it’s useful | |---|---|---| | Adapter | One MCP server at /mcp that aggregates stdio MCP, OpenAPI, and manual HTTP tools | Turn “many backends” into one MCP endpoint (runs well inside a trusted network) | | Gateway | Public-facing MCP endpoint /{profile_id}/mcp that proxies/aggregates upstream MCP sources and applies auth/policy | Run multi-tenant, policy-driven “virtual MCP servers” without exposing upstreams | | Tenant | Isolation boundary in Mode 3 (Postgres) | Separate teams/projects/environments with isolated secrets and API keys | | Profile | One “virtual MCP server” exposed as /{profile_id}/mcp | Create focused tool surfaces per use case (dev vs prod, read-only vs write, etc.) |

What you can build with it

• Expose a REST API as MCP tools:
• OpenAPI → tools, or
• manual HTTP tools DSL → tools
• Publish stdio MCP servers over HTTP by spawning them as child processes and re-exposing the tool surface at /mcp.
• Aggregate multiple tool sources into one MCP endpoint (and keep names stable by collision prefixing when needed).
• Operate a tenant-scoped gateway:
• API keys and optional OIDC/JWT for data-plane auth
• per-profile tool allowlists, transforms (rename/defaults/parameter tuning), and tool-call limits (timeouts/retries/quotas)
• HA-friendly session routing via Gateway session tokens (Mcp-Session-Id)

Deployment configurations

1. Solo developer

Just the Gateway, no Adapters needed.

Use the Gateway with gateway-native HTTP/OpenAPI sources and connect your MCP client to a single profile endpoint.

flowchart LR
  Client["MCP client"] -->|"streamable HTTP '/{profile_id}/mcp'"| Gateway["Gateway"]
  Gateway -->|"gateway-native HTTP tools"| HttpApi["Your HTTP API"]
  Gateway -->|"gateway-native OpenAPI tools"| OpenApi["OpenAPI (spec or API)"]

2. Team with multiple projects

Gateway with Tenants and Profiles.

Each tenant owns multiple profiles, where each profile is a focused “virtual MCP server” with its own tool surface and data-plane auth/policy.

flowchart LR
  TeamA["Team A MCP clients"] -->|"profile A"| Gateway["Gateway"]
  TeamB["Team B MCP clients"] -->|"profile B"| Gateway

  Gateway --> TenantA["Tenant: team-a"]
  Gateway --> TenantB["Tenant: team-b"]

  TenantA --> ProfileDev["Profile: dev-tools"]
  TenantA --> ProfileProd["Profile: prod-readonly"]
  TenantB --> ProfileDS["Profile: data-science"]

  ProfileDev --> InternalApis["Internal APIs"]
  ProfileProd --> ProdApis["Prod APIs (GET only)"]
  ProfileDS --> MlApis["ML APIs + data tools"]

3. Complex / company-wide

Gateway + Adapters for aggregation and splitting.

Use Adapters to bundle related systems into upstream MCP endpoints (typically inside a private network), then use Gateway profiles to expose focused tool surfaces to clients.

flowchart LR
  Client2["MCP client"] -->|"streamable HTTP '/{profile_id}/mcp'"| Gateway2["Gateway"]

  Gateway2 -->|"proxy to upstream MCP"| AdapterA["Adapter A (/mcp)"]
  Gateway2 -->|"proxy to upstream MCP"| AdapterB["Adapter B (/mcp)"]

  AdapterA -->|"aggregates"| Billing["Billing API"]
  AdapterA -->|"aggregates"| Inventory["Inventory API"]
  AdapterA -->|"aggregates"| AuthSvc["Auth service"]

  AdapterB -->|"aggregates"| Mlflow["MLflow"]
  AdapterB -->|"aggregates"| Jupyter["Jupyter"]

Adapter: what it exposes

The Adapter exposes one MCP server over streamable HTTP:

• MCP endpoint: /mcp
• Operational endpoints: /health, /health/any, /health/all, /ready, /status, /map

Documentation

• Start here (workspace docs index): [docs/INDEX.md](docs/INDEX.md)
• Component docs:
• Adapter: [docs/adapter/INDEX.md](docs/adapter/INDEX.md)
• Gateway: [docs/gateway/INDEX.md](docs/gateway/INDEX.md)
• Gateway CLI: [docs/gateway-cli/INDEX.md](docs/gateway-cli/INDEX.md)
• Web UI: [docs/ui/INDEX.md](docs/ui/INDEX.md)
• CI/CD + releases: [docs/CICD.md](docs/CICD.md)
• Workspace layout: [docs/WORKSPACE.md](docs/WORKSPACE.md)

Other ways to run (local dev)

This section shows a few common ways to run pieces locally. Many commands have convenient Make targets (see make help).

Local (cargo)

• Build + run the adapter with the example config (recommended):
• make adapter-run
• Or build a static release binary:
• make build-release-adapter

Equivalent raw cargo commands:

• cargo build --release -p unrelated-mcp-adapter
• cargo run -p unrelated-mcp-adapter --bin unrelated-mcp-adapter -- --config ./tests/fixtures/test-config.yaml --bind 127.0.0.1:8080

• More example configs (and a handy playground) live under [tests/fixtures/](./tests/fixtures/) — these are used by the demo stack and integration tests, and some are intentionally invalid to demonstrate validation errors.

Docker

• Build the adapter image:
• make docker-build-adapter
• Run the adapter container (standalone, without compose):
• make docker-run-adapter

Docker Compose (Gateway + multiple Adapters)

The default docker-compose.yml now spins up:

• Postgres for Mode 3 (gateway_db) + a migrator job (gateway_db_migrate)
• a Mode 3 Gateway (gateway) exposing:
• data plane: http://127.0.0.1:27100
• admin/control plane: http://127.0.0.1:27101 (token: dev-admin-token in compose)
• a Web UI (gateway_ui) on http://127.0.0.1:27102
• 3 adapters (adapter_http_tools, adapter_openapi, adapter_stdio_aggregation) plus example backends (httpbin, petstore)

Run:

• make up

Then create a tenant + profile (preferably via the Web UI) and connect your MCP client to:

• Gateway data plane: http://127.0.0.1:27100//mcp

Notes:

• The compose stack sets a dev-only default UNRELATED_GATEWAY_SECRET_KEYS so Mode 3 tenant secrets can be encrypted at rest (override this in real deployments).
• Adapters are still exposed individually on their mapped ports for debugging.

Gateway admin CLI:

• Run: make cli-dev CLI_ARGS="--help" (docs: [docs/gateway-cli/INDEX.md](docs/gateway-cli/INDEX.md))

Web UI:

• Open http://127.0.0.1:27102
• On a fresh install (bootstrap enabled + empty DB), onboarding will guide creating the first tenant.

Install options

• Docker image: ghcr.io/unrelated-ai/mcp-adapter
• Stable tags publish :latest + :X.Y.Z
• Pre-release tags publish :X.Y.Z-rc.N
• GitHub Release binaries: static Linux build (x86_64-unknown-linux-musl)

If your stdio MCP servers require extra runtimes (node/python/etc), build your own image and copy the adapter binary from the published image:

FROM alpine:3.20
RUN apk add --no-cache ca-certificates nodejs npm
COPY --from=ghcr.io/unrelated-ai/mcp-adapter:latest /app/unrelated-mcp-adapter /app/unrelated-mcp-adapter
ENTRYPOINT ["/app/unrelated-mcp-adapter"]

The published adapter image is a minimal image containing a static adapter binary, so it can be copied into Alpine/Debian/etc (same CPU architecture).

Examples

Each example below is a complete YAML config file (you can also use JSON). The key part is the servers: map.

Once running, the adapter serves the resulting MCP tools over streamable HTTP (/mcp). Tool names are usually the tool key (like create_invoice); if two servers define the same tool name, the adapter prefixes on collision (serverName:toolName).

HTTP tools (manual HTTP DSL)

servers:
  billing_api:
    type: http
    baseUrl: http://billing-api:8080
    tools:
      create_invoice:
        method: POST
        path: /v1/invoices/{customerId}
        description: Create an invoice for a customer and return the created invoice JSON.
        params:
          customerId:
            in: path
            required: true
            schema:
              type: string
          dryRun:
            in: query
            schema:
              type: boolean
          body:
            in: body
            required: true
            schema:
              type: object
              properties:
                amountCents:
                  type: integer
                currency:
                  type: string
              required:
                - amountCents
                - currency

Result: exposes an MCP tool named create_invoice (or billing_api:create_invoice on collision). It takes customerId (string, path), dryRun (boolean, query), and body (JSON object).

OpenAPI (auto-discover tools from a spec)

servers:
  orders_api:
    type: openapi
    spec: ./openapi/orders.yaml
    baseUrl: https://api.example.com
    autoDiscover:
      include:
        - "GET *"
      exclude:
        - "DELETE *"

Result: exposes one MCP tool per discovered OpenAPI operation (here: all GET operations except DELETE). Tool names come from operationId when present; tool parameters come from the OpenAPI schema (path/query/header/body).

MCP stdio (spawn a local MCP server process)

servers:
  local_mcp:
    type: stdio
    command: /usr/local/bin/my-mcp-server
    args:
      - "--mode"
      - "prod"
    env:
      API_TOKEN: "${API_TOKEN}"

Result: spawns the process and re-exposes whatever tools/prompts/resources that MCP server provides (through the adapter’s MCP endpoint, /mcp). env values support ${VAR} expansion.

Project meta

• Changelog: [CHANGELOG.md](CHANGELOG.md)
• Contributing: [CONTRIBUTING.md](CONTRIBUTING.md)
• Security policy: [SECURITY.md](SECURITY.md)

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: unrelated-ai
• Source: unrelated-ai/mcp-gateway
• License: MIT

Install and usage instructions live in the source repository linked above.