Trust Gate MCP

Tamper-evident receipts for consequential agent actions, as an MCP server, with a decision gate that returns ALLOW, DENY or ESCALATE.

Seven tools, one shared signing primitive: the open-source OpenAgentOntology mint_receipt.

What actually gets signed depends on what you install -- OAO detects its backend at import time, so this is worth stating plainly rather than advertising the best case:

PQ-required verify (the default) demands at least one verified post-quantum leg. That shows a post-quantum signature is present and valid over the same bytes. It does not say whose key made it (see "What a valid receipt proves"), and it does not make the default backend production-grade.

What the gate does and does not do

• Scope. It applies only to actions your code routes through gate_decision. It does not observe, intercept or block anything an agent does, and it does not check afterward whether an action happened. Your code (a hook, a wrapper) must ask first and must obey the answer.
• ALLOW is a closed list. An action gets ALLOW only when its name is a known read verb (read, get, list, search, query, find, show, view, inspect, describe) followed only by words from a small built-in vocabulary of nouns (for example read_file, list_files, search_files, list_dir, get_user, describe_table), the name is a plain identifier, and the resource is built only from words of that vocabulary, a few ordinary document words (readme, guide, changelog, ...) and source or prose file types (md, txt, py, ...), with digits allowed between words (docs/2024/notes.md). A name that contains any other word escalates to a human, because the gate has not seen it. This is deliberately narrow: most real file names escalate. Widening it means adding words to the vocabulary in the source, which is a visible change.
• Structural rules escalate as well. A hidden path (.env, .ssh, .git, ..), an absolute path, a URL, a host name as the first part of a path (example.com/x, localhost), a network address in any of its usual spellings (four dotted numbers of any width, hexadecimal, octal such as 0177.1, one long integer, 0; short dotted numbers such as 1.2.3 stay versions unless the last number is too large for a version or the first is 0 or 127), a resource that starts with -, a phrase of three or more words, anything containing characters other than letters, digits, spaces and _ . - / (after Unicode compatibility folding), and a file whose real type is configuration, data, key, state or archive (a rotation counter such as .1 is ignored when judging the type; a listed data or archive type earlier in a dotted name, such as users.csv.txt, counts too). A resource that carries a command escalates: a listed mutating word (alter, merge, flush, load, ...), a critical verb glued to a word (deleteall), or any multi-word resource after a query verb.
• DENY is for a critical lead verb (payments, deletion, deployment, code execution, privilege changes). A risky word elsewhere in the name, or in the resource, escalates so that a human decides; so do the high-risk lead verbs (format, remove, reset, release, escalate) and a file name that contains a risky word (test_delete_user.py).
• Risky and sensitive words are listed, and the lists are finite. Listed secret and regulated-data words (password, token, key, pass, config, ssn, diagnosis, ... in their glued, plural and digit-swapped forms) give the HIGH tier a specific reason. The lists are not what keeps a name from being allowed (a sensitive name on no list still escalates, because ALLOW is closed). history is not a read noun. Matching runs across word boundaries, so some harmless names show the sensitive reason (api/index.md contains "pii"). A lookup such as list_users is a read like any other: the gate does not judge how much personal data it returns. Common read tools whose names contain a word that is also a risky verb (get_email, list_messages) escalate.
• It is a name-based heuristic, not a proof. ALLOW does not mean the action is safe: an agent that calls a dangerous tool read_file is not caught, because the gate sees names, not behaviour. context is recorded in the receipt as a hash and is not used to decide. There is no operator override in this package.
• Read the verdict from the receipt, and check who issued it. receipt.decision (and receipt.evidence.ontology.verdict) are signed. The top-level verdict and permit fields of the response are not; verify_receipt lists the unsigned keys of any receipt it checks. A receipt is a gate verdict from your server only if verify_receipt with expected_kid set to your server's kid returns signer_pinned: true AND verified.issuer_tool is gate_decision AND verified.operation is decision_gate_commit; every other tool signs its own issuer_tool, and without the pin those fields are only claims made by whoever signed.
• No permit without a receipt this server would accept. COMMIT issues GRANTED only if the receipt was minted and signed, and verifies under this server's own policy (in PQ-required mode that includes a valid post-quantum signature).
• Action and resource are recorded in clear text in the signed receipt. Do not put secrets or personal data in them.
• A receipt is evidence of a verdict, not a log entry. atom_id is derived from the content (40 bits of the evidence hash), so identical decisions share one and different decisions can too; key on evidence_hash and signed_at. The server keeps no log of the receipts it signs: store the ones you need. There is no replay or freshness check: compare evidence.ontology.action and resource with what you are about to do.
• 0.2.1 returned GRANTED for every action. Upgrade to 0.3.0. See the changelog.

What a valid receipt proves

verify_receipt returning ok: true means the evidence is unchanged since it was signed, the receipt carries a signature, any kid it carries matches the key it carries, and (in the default PQ-required mode) at least one post-quantum signature verified over the same bytes.

It does not mean the receipt came from a server you trust: anyone can generate keys and sign a receipt. To check who signed, pass expected_kid (the kid of the Ed25519 key you trust, learned out of band; 32 hexadecimal characters, case and surrounding spaces ignored, a blank or malformed pin is refused). A pin counts only when that key's own Ed25519 signature verifies, and signer_pinned: true says it did (a signature verifying shows that the key signed those bytes, not that anyone holds its private key: Ed25519 accepts a few special public keys that verify for any message, which cannot affect a pin on a key you trust); ok: true with signer_pinned: false is an integrity result only, and kid is reported only when the Ed25519 signature verified. A receipt that names an Ed25519 signer (key or kid) without a valid signature from it is refused: those two values are public and can be copied into a receipt signed with other keys.

On stdio the signing key is stored unencrypted under the user's home directory, so an agent running as that user can read it: a pin does not protect against the agent you are gating. The kid fingerprints the Ed25519 key only. The ML-DSA and hash-based keys inside a receipt are not covered by it, so a pinned check is only as strong against a quantum attacker as Ed25519. Results can also depend on which post-quantum backends the verifying host has installed (a leg it cannot check is reported unverifiable).

Hardening

• H1 key persistence + bootstrap with FAIL-CLOSED kid-drift check (container deploy; the bootstrap checks the Ed25519 key only)
• H2 per-IP token-bucket rate limit (FIFO eviction). It parses the request body exactly as the MCP transport does, from the raw bytes (a byte-order mark, UTF-16 and UTF-32 are read too), and charges the tool it names; the signing tools share the signing budget. A body it cannot parse, or one over 64 KiB, is charged to the signing budget; a request is refused with 413 as soon as its body exceeds 1 MiB (at most 1 MiB is buffered). Signing (about 70 ms per call with the default backend) and verification (about 20 ms) run on the event loop and the budgets are per client, so there is no overall cap on that work. Clients are identified by the connection address as uvicorn resolves it: set TRUST_GATE_FORWARDED_ALLOW_IPS to your proxy's address, or any client can choose its own identity with an X-Forwarded-For header.
• H3 PQ-required verify: refuses a receipt with no verified post-quantum signature (and any unsigned receipt)
• H4 128-bit kid on every minted receipt, recomputed from the canonical key bytes and checked on verify
• Optional bearer auth and a narrowed CORS allowlist via TRUST_GATE_BEARER_TOKEN + TRUST_GATE_ALLOWED_ORIGINS
• The test suite pins every listed verb, noun and sensitive word individually (from an independent copy of the lists) and tests their inflected, plural, glued, digit-swapped and compatibility-character forms, plus adversarial cases for receipt forgery and PQ stripping; real MCP requests are posted through the whole HTTP stack

Deploying over HTTP

The HTTP entrypoint (server_http.py) is written for the container image, where its modules sit side by side; with a pip install, use the stdio server. Bearer auth is off by default and CORS is * in that mode. Anyone who can reach an open deployment can obtain receipts signed with its key, including "ALLOW" receipts for read-only names, and a receipt from an open deployment says nothing about who asked. If you host it, set TRUST_GATE_BEARER_TOKEN and TRUST_GATE_ALLOWED_ORIGINS.

The HTTP entrypoint counts ?via=<label> visits as one log line (label, kind, user-agent family). Only three built-in labels plus those named in TRUST_GATE_CHANNELS (comma separated) are recorded; the host's own access log also records request lines, including client addresses.

DNS-rebinding protection is off and there is no switch for it: on a local HTTP server without bearer auth, any web page you visit can call it. Put it behind a proxy that checks the Host header, or set TRUST_GATE_BEARER_TOKEN.

Install (stdio)

pip install trust-gate-mcp
trust-gate-mcp

Add [slh] for the native ML-DSA backend (see the table above). From a checkout, pip install -e ".[dev]" then python -m trust_gate_mcp.

Container deploy (Smithery / any container host)

docker build -t trust-gate-mcp .
docker run -p 8081:8081 -v trust-gate-data:/data/oao trust-gate-mcp

The volume mount on /data/oao is required for production: without it the signing key changes on every container restart, receipts signed before a restart still verify from their own certificate, but a pinned kid stops matching new ones. The persistent key_metadata.json holds the signer's kid; the bootstrap step refuses to start if it drifts. The image installs dilithium-py by default; see the table above.

License

Apache-2.0 (full text in LICENSE). Copyright 2026 Cyber Warrior Network. Built on the open-source OpenAgentOntology primitive.