Sovereign Gateway

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.

[Built on patent-pending components]

pip install sovereign-mcp-gatewaysovereign-mcp-gateway --init          # writes gateway.json from the servers you already runsovereign-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-gitpython 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 checkupstreams: 2layers:   policy -> intent -> text-filter -> frozen-verify -> audit
EXPOSED AS                             UPSTREAM TOOLgit__git_status                        git.git_statusgit__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 extrapip install "sovereign-mcp-gateway[text,intent]"      # severalpip 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                                  # baselayers:   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:

LAYER C  - probing the configured models with one real call--------------------------------------------------------------  OK. The configured models produced identical documents.  Layer C will pass ordinary output rather than refusing it.

Consensus compares canonical hashes, so two models that are both semantically right but structurally different never agree. A weaker model that echoes the schema back —

{"branch": {"type": "string", "value": "main"}}   instead of   {"branch": "main"}

— mismatches on every call, forever, and the gateway refuses everything with a reason that correctly reads "the models disagreed". Because they did.

The probe distinguishes the three outcomes:

Install sovereign-mcp-gateway[consensus] or [all] — the HTTP providers need requests, which the core library deliberately does not depend on.

--check also lists the active layers, so you can confirm at a glance:

layers:   policy -> intent -> text-filter -> frozen-verify -> consensus -> audit

If consensus is absent from that line, it is not running, whatever the config says.

Namespacing

With namespace on (the default) a tool is exposed as git__git_status. Two upstreams offering the same tool name cannot collide, shadow each other, or be reached through the wrong namespace. Turn it off only when you have a single upstream.

Policy

"policy": {  "deny_tools":  ["git__git_reset", "write_query"],  "allow_tools": null,  "pii_policy":  "warn",  "fail_closed": true,  "rate_limit_interval": 0}

• deny_tools matches either the exposed name (git__git_reset) or the upstream tool name (git_reset, on every upstream that has it).
• allow_tools, when set, refuses everything not listed.
• pii_policy defaults to warn, not block. Real tools return personal data as normal output — every git log entry carries an author email — and blocking those makes the gateway unusable. Set block when your tools should never emit PII.
• fail_closed decides what happens when a layer itself errors. Default: refuse.
• rate_limit_interval is 0, which disables the behavioural floor's own inter-action delay. That delay is right for one agent taking deliberate steps and wrong for a proxy, where a burst of tool calls is ordinary traffic.
• entropy_policy defaults to warn. The text filter's entropy heuristic hunts for encoded payloads hidden in prose, but tool arguments are routinely structured — paths, identifiers, hashes — where high entropy is normal. A temporary directory path alone was enough to have a legitimate call refused. Set block when your arguments really are prose.

What this does not do

It verifies calls against frozen definitions and inspects arguments and results. It does not read your servers' source, so it cannot see a check that is present, is called, and silently does nothing. That still takes someone reading the implementation.

It also cannot protect against a compromised upstream returning correct-looking data — Layer C consensus in sovereign-mcp addresses that, and requires model providers you configure yourself.

Licence

Business Source License 1.1 — see LICENSE.

The source is public. You may read it, modify it, create derivative works and use it for development, evaluation and any other non-production purpose at no cost.

Production use is also free for an individual, or an organisation of four or fewer people — that is written into the licence as an Additional Use Grant, not just stated here. Larger organisations need a commercial licence.

Each version converts to Apache 2.0 on its Change Date, four years after publication.

To license it for production, or to ask whether your use needs one: [email protected]