sovereign-mcp-gateway

A gating proxy for Model Context Protocol servers. Point your MCP client at the gateway instead of at your servers. It connects to every upstream you list, merges their tool catalogues into one, and puts every call through a verification chain before it reaches the server that would execute it.

pip install sovereign-mcp-gateway
sovereign-mcp-gateway --init          # writes gateway.json from the servers you already run
sovereign-mcp-gateway --config gateway.json --check

--init reads the MCP configuration you already have (Claude Desktop, Claude Code, Cursor, VS Code or Windsurf) and writes a gateway.json that proxies those same servers, so the first run produces a working config rather than a configuration error. It will not import the gateway's own entry, because that would make it proxy itself.

The gateway is itself an MCP server, so any client that speaks MCP works with no changes.

That base install is a working gateway. Four optional extras add further layers on top — see Installing.

What it stops

An agent reads a GitHub issue whose body carries an instruction aimed at the model rather than at you. It is persuaded, and calls git_commit.

Same tool, same arguments, same server. The difference is whether anything was in a position to refuse.

Read the walkthrough: Your agent reads an issue — or run it:

pip install "sovereign-mcp-gateway[all]" mcp-server-git
python examples/poisoned_issue.py

Why a proxy and not a library

A library has to be adopted by whoever wrote the server. A proxy protects servers you cannot modify — which is most of them, because the useful MCP servers are published packages someone else maintains.

It also gives you one place to hold policy and one audit trail across every server an agent can reach, rather than per-server configuration nobody keeps in sync.

Configure

Start from what you already run

$ sovereign-mcp-gateway --init

Wrote gateway.json

  imported 3 servers from Claude Desktop
    /Users/you/Library/Application Support/Claude/claude_desktop_config.json
  imported 1 server from VS Code (project)

  upstreams: fetch, git, sqlite, time

  skipped:
    sovereign - this gateway - importing it would proxy itself
    notion    - no command, probably a remote/SSE server
    git       - already imported from another client

Four things it will not do: import itself, import a remote server it cannot launch as a subprocess, overwrite an existing file without --force, or write a deny_tools list you did not choose. It writes the file, tells you what it took and what it left, and stops.

Pass --config PATH alongside --init to write somewhere other than ./gateway.json.

After running it, replace those servers in your client with a single entry for the gateway. Leaving both means your agent talks to them directly as well as through the proxy, and the audit trail will only show half the traffic.

{
  "servers": {
    "git":    {"command": "mcp-server-git",    "args": ["--repository", "/repo"]},
    "sqlite": {"command": "mcp-server-sqlite", "args": ["--db-path", "/data.db"]}
  },
  "policy": {"deny_tools": ["git__git_reset"], "pii_policy": "warn"},
  "audit":  {"path": "gateway-audit.jsonl"}
}

Check the wiring before a client ever sees it:

sovereign-mcp-gateway --config gateway.json --check

SOVEREIGN GATEWAY - configuration check
upstreams: 2
layers:   policy -> intent -> text-filter -> frozen-verify -> audit

EXPOSED AS                             UPSTREAM TOOL
git__git_status                        git.git_status
git__git_reset                         git.git_reset          [DENIED]
sqlite__read_query                     sqlite.read_query
...
18 tools exposed.

The chain

policy → intent → text-filter → frozen-verify → [ call executes ] → output-verify → logic-rules → audit

Installing

The base install is a working gateway, not a stub:

pip install sovereign-mcp-gateway

That gives you policy → frozen-verify → audit, which already refuses a tool no upstream exposes, an argument of the wrong type, an undeclared parameter, a tool on your deny list, and prompt injection in an argument. Nothing else needed.

Each extra adds a layer on top:

Combine what you want, or take everything:

pip install "sovereign-mcp-gateway[text]"             # one extra
pip install "sovereign-mcp-gateway[text,intent]"      # several
pip install "sovereign-mcp-gateway[all]"              # every layer

All four extras are small pure-Python packages — [all] adds no compiled dependencies and no service to run.

A partial install degrades visibly. The gateway prints its active layers at startup, so you can always see what is actually running:

layers:   policy -> frozen-verify -> audit                                  # base
layers:   policy -> intent -> text-filter -> frozen-verify -> audit         # [all]

If a layer isn't on that line, it isn't running — whatever you think you installed.

Verified end to end

Against mcp-server-git and mcp-server-sqlite running as real upstreams, driven by a real MCP client:

Afterwards the repository still holds one commit and the database holds exactly the row it should — checked by opening them directly, not by trusting the gateway's own report. Eleven audit records for ten calls; editing any one of them breaks the chain.

Those cases are the test suite, not a screenshot: pytest tests/ -v.

Layer C: N-model consensus

Every other layer is deterministic and local. Layer C is the exception: it asks several independent models to extract the same structured document from a tool's result, canonicalises each answer, and compares the SHA-256 hashes. Agreement is decided by hash, not by prose.

It is off unless configured, because it is the only layer that costs money and latency per call, and the only one that sends tool output to a model.

{
  "servers": { "...": {} },
  "consensus": {
    "providers": [
      {"type": "local", "model": "llama3.1:8b"},
      {"type": "local", "model": "qwen2.5:7b", "base_url": "http://localhost:11434/v1"},
      {"type": "openrouter", "model": "anthropic/claude-3.5-sonnet",
       "api_key_env": "OPENROUTER_API_KEY"}
    ]
  }
}

Two provider types: local (any OpenAI-compatible endpoint — Ollama, vLLM, LM Studio; base_url defaults to http://localhost:11434/v1) and openrouter (the key is read from the named environment variable, never written in the config).

Three rules the gateway enforces at startup rather than discovering at runtime:

• At least two providers. One model cannot disagree with itself; a consensus of one reports agreement on every call, which is worse than no layer because it looks like verification.
• No duplicate models. Two instances of the same model agreeing is not independent verification.
• A missing API key refuses to start. It does not fall back to running without the layer.

All providers run at temperature = 0, enforced in the constructor.

Check your models agree before you trust the layer

--check runs one real consensus call against your configured models and tells you what happened. This matters more than it sounds: