Gargoyle

Signed, single-use, argument-bound receipts and human-approval holds for agent tool calls.

An agent asks for permission to run a tool. Something you already trust decides (OPA, Cedar, your own code). Gargoyle turns that decision into a receipt: an Ed25519-signed record that names the tool, a hash of its arguments, the principal the decision was made for, the target it may be presented to, and the policy that decided. The target checks the receipt before it acts, and honors it once. When the policy says a call needs a human, the receipt is held until an approver signs a grant for exactly that decision with a key the issuing service does not have.

Gargoyle does not decide whether a call is allowed. It makes an allow decision portable, verifiable, bound to one call by one caller, and spendable once. The decision comes from your policy engine (an OPA adapter is included); the caller is authenticated by your transport (the proxy-verified client certificate, by default); Gargoyle sits in front of your MCP server and lets through only what both agreed to.

Status: pre-release (0.6.1.dev0, on PyPI as gargoyle-receipts). The wire format is gargoyle/3 and may still change before the first release; it is specified in SPEC.md.

What a receipt gives you

• Binding. A receipt for transfer_funds with {"amount": 250} is refused for {"amount": 250000}, for delete_database, and at any target but the one it names.
• A principal. The receipt carries the identity the issuer authenticated; the target compares it with the identity it authenticated. A receipt obtained by one agent cannot be spent by another. When the agent acts for a user, the receipt records that too (on_behalf_of).
• Standard encodings. A receipt can travel as a JWS (EdDSA) that any JWT library verifies, and the issuer can take the caller's identity from an OAuth bearer token, an RFC 8693 delegation token or a Transaction Token, signed EdDSA, RS256 or ES256 by keys from the provider's JWKS. With Entra ID the principal is the user's UPN and the token's groups and app roles become policy inputs, so a rules file can say "finance may hold invoice updates; shipping may not" without naming people.
• Single use. Every receipt carries a signed decision_id; an executor claims it exactly once. A replayed receipt is refused.
• Expiry. Every receipt carries a signed not_after (default 300 s). A captured receipt, or a leaked issuer key, is useful for minutes, not forever.
• Currency. The receipt pins the policy that decided. If the policy changes, outstanding receipts stop being honored and the call must be decided again.
• Holds. A decision that requires approval is flagged inside the signed receipt. The target will not run it without a grant a human signed for that exact decision, under their own key, on their own device.
• Proof of execution. An executor with a key answers each honored call with a signed execution receipt: the decision it fulfilled, the policy version that allowed it, and a hash of the outputs. Receipt, grant and execution receipt verify together, offline, with three public keys.

Install

pip install -e ".[dev]"
python -m pytest

Python 3.12 or later. A TypeScript verifier and consumer for Node 20+ lives in ts/ (cd ts && npm install && npm test); its suite recomputes the same conformance vectors byte for byte, so the two implementations cannot drift apart silently. Redis is optional (.[redis]) and only needed to share the replay cache and pending-approval set across executor replicas.

Issuing a receipt

The policy is anything with pin() and decide(); see gargoyle.policy. The interaction you pass carries the call (tool, args, principal) and whatever claims your policy reads. The receipt records only the call and the policy pin.

from gargoyle.issuer import ENVELOPE_HEADER, issue

who = principal_from_token(bearer_or_transaction_token, IDP_KEYS, audience="issuer")  # gargoyle.identity
# IDP_KEYS: pinned, or entra_keys("<tenant>") / keys_from_jwks(<provider JWKS>)
interaction = {
    "tool": "transfer_funds",
    "args": {"amount": 250, "to": "acct-9"},
    "principal": who["principal"],         # the agent, as the identity provider authenticated it
    "on_behalf_of": who["on_behalf_of"],   # the user it acts for, when the token says so
    **claims,                              # whatever your policy decides on
}
result = issue(policy, "https://tools.example.com/target", interaction,
               signing_key=issuer_private_key, key_id="issuer-2026-10")

if result.envelope is None:
    refuse(result.decision.reason)
elif result.decision.requires_approval:
    hold(result.envelope)            # see "Approval holds"
else:
    send(call, headers={ENVELOPE_HEADER: json.dumps(result.envelope)})
    # or, for a target that speaks JWT: gargoyle.jws.encode_receipt(result.envelope, issuer_private_key)

Open Policy Agent. gargoyle.opa.OpaPolicy asks OPA's data API and pins the exact set of policy modules OPA holds, so editing a policy invalidates every outstanding receipt until the call is decided again.

from gargoyle.opa import OpaPolicy
policy = OpaPolicy("http://127.0.0.1:8181", "gargoyle/decision")

package gargoyle
default decision := {"allow": false, "reason": "DEFAULT_DENY"}
decision := {"allow": true, "requires_approval": input.tool == "transfer_funds"} {
    input.principal == "spiffe://example.org/agents/payments"
    input.tool in {"read_balance", "transfer_funds"}
}

OPA unreachable, an undefined document, or a result in no recognized shape is a refusal, never an issuance. Any other engine goes behind the same two-method interface.

No engine. gargoyle.rules.RulesPolicy decides from a JSON file: first matching rule wins, globs on tool name, principal, delegation, the caller's groups and roles, and string arguments, actions allow, hold or deny. The pin is the file's hash, so an edit invalidates outstanding receipts exactly as an engine change does. This is what gargoyle init sets up.

Checking a receipt

Every executor runs the same check: establish the live policy pin, establish the trusted issuer keys, verify the receipt against the call it is about to run, claim the decision_id. gargoyle.consumer.Consumer packages it.

from gargoyle.consumer import Consumer

consumer = Consumer(
    target_id="https://tools.example.com/target",
    pinned_public_keys={"issuer-2026-10": issuer_public_key},
    record_bytes=open("published_record.json", "rb").read(),   # {policy_id, policy_sha256}
)

call = {"tool": name, "args": arguments, "principal": peer_identity}
verdict = consumer.check(envelope, call)
if verdict.honored:
    run(name, arguments)
else:
    refuse(verdict.reason)           # REF_VERIFY_* or REF_TARGET_*

peer_identity is who you authenticated as the caller (an mTLS peer, a bearer token's subject, an identity a trusted proxy asserted). It is never read from the request. A consumer refuses to be built without issuer-key trust, and the verifier refuses to run without it, so an unsigned receipt is never honored. For a flagged receipt, pass the human's grant as grant=; the consumer verifies it against the receipt's own decision and hold, claims it once, and only then claims the receipt.

For a live pin that can be reissued, give the consumer a record source (a publisher-signed record with a serial and an expiry) instead of a static pin; for issuer keys that can be rotated and revoked, a key record source. Both are consulted on every check and fail closed.

Proof that it ran

Give a consumer an executor key and it signs an execution receipt after the side effect. Anyone holding the issuer's, approver's and executor's public keys can then audit one call with no network and no live state:

from gargoyle.execution import verify_chain

chain = verify_chain(receipt, execution,
                     issuer_public_keys=ISSUER_KEYS, executor_public_keys=EXECUTOR_KEYS,
                     grant=grant, approver_public_keys=APPROVER_KEYS,
                     outputs=outputs, check_outputs=True)
chain["accepted"], chain["steps"]   # True, [("authorization", ...), ("approval", ...), ("execution", ...)]

That is the sentence Gargoyle can say that nobody else can: a specific policy version authorized a specific call for a specific principal, a human's key released it, and exactly that call ran once, verifiable offline by anyone, in Python or TypeScript.

The MCP middleware also keeps an evidence log: one hash-chained line per decision, refusals and declined holds included, so the trail answers "what did the agent try" as well as "what did it do". gargoyle audit --log verifies every entry and the chain, and prints the tail hash to note down so a later copy that lost its end is caught (SPEC.md section 15).

Where it sits

MCP gateways (TrueFoundry, Monosign, hoop.dev, Akto and others) hold a tool call, notify a reviewer and log the click. Gargoyle is not a gateway. It runs beside the client, and its approval is a signature made with a key the issuer never holds, checked at execution and again by anyone later. In 2026 three Internet-Drafts describe parts of the same shape: Acta signed receipts (an access decision, signed), EP authorization receipts (a human's pre-execution signoff), XAIP receipts (a signed execution record). Gargoyle chains all three on one decision id with the policy version inside, uses the same RFC 8785 canonical form they do, and SPEC.md section 16 maps every field. A hold can also be announced to Slack or any webhook (GARGOYLE_HOLD_WEBHOOK_URL); the notice carries no key, so nothing in a chat channel can approve anything.

Deploy it

pip install gargoyle-receipts

Three commands, no policy engine, no containers, no client changes:

gargoyle init /path/the/agent/may/touch     # keys, config, rules.json
gargoyle issuer                             # the issuing service, on rules.json
gargoyle approve --watch 2                  # where the approver sits (terminal)

or approve in a browser: open the issuer's /approve page and load approver-web.json, or, with Entra configured, sign in with Microsoft; the page shows each held call with its full arguments, checked against the receipt, and signs the grant locally with WebCrypto. The key never leaves the browser, and with Entra it is generated there, non-exportable, and enrolled under the approver's UPN, so grants and audits name the person. Then paste the server entry from gargoyle-deploy/mcp-config.json into your MCP client. docs/guide-hyperv-ubuntu-apache.md is the full single-instance build: Ubuntu VM on Hyper-V, Apache with TLS, a Windows 11 client running Claude Code. Reads are free, writes block until the approver's key signs, deletes are refused; the policy is a JSON rules file you edit. A hold can demand more than one signature: "approvals": 2 on a rule, with an optional "approvers": [...] naming which keys count, puts the terms in the receipt itself, and gargoyle init --approvers 2 creates the keys. Every honored call comes back with receipt, grants and execution receipt, and gargoyle audit verifies them later offline. examples/filesystem-agent walks through it, and through the alternative with OPA in Docker for sites that run a policy engine.

The shipped consumer

MCP middleware (python -m gargoyle.mcp_server -- <upstream command>): a stdio MCP server you put in front of any other stdio MCP server. The handshake, tools/list and everything else are relayed. A tools/call is forwarded only when params._meta["gargoyle/envelope"] carries a valid receipt for that tool, those arguments and the principal the server serves (and params._meta["gargoyle/grant"] a grant, when the receipt is flagged); a refused call is answered here and never reaches the upstream. Without an upstream command it runs a built-in demo executor. With GARGOYLE_MCP_ISSUER_URL and GARGOYLE_MCP_CLIENT_TOKEN it obtains a receipt from the issuing service for any call that carries none, and waits for the human's grant when the call is flagged, so clients that cannot attach receipts work unchanged. Configuration is by environment:

Issuing service (gargoyle.issuer_service, run with uvicorn from app_from_env()): identity token in, receipt out, with the hold and grant exchange for flagged calls (/issue, /pending, /grants, /decline, /record, /keys) and the browser approval console at /approve. Policy from a rules file or OPA; identity from a pinned key or, with GARGOYLE_ENTRA_TENANT, the tenant's signing keys (both at once when a pinned key is also set, so init-minted approver tokens keep working while users move to gargoyle login).

Approval holds

When the policy's decision carries requires_approval, the receipt is issued with requires_approval: true and a signed approval_request_id. No target will run it on its own: the consumer demands a grant for exactly that decision and hold, verifies it (provenance, separation of duties, binding, expiry), claims it once, and only then claims the receipt. A human runs gargoyle-approver on a machine that holds the approver private key; given the receipt it shows the tool, argument hash, principal, target and policy that are inside the issuer's signature, and emits the grant. The agent presents receipt and grant together. Approver keys should come from a signed key record where the key's role is approver (gargoyle.approver_trust), so an issuer-controlled key can never approve. An issuer that prefers to hold flagged receipts until the grant exists can do that too (gargoyle.pending_store); the target's check does not depend on it.

Modules

What Gargoyle does not do

• It does not decide. The policy does, and the receipt is only as good as that decision.
• It does not authenticate principals. The issuer and the target each authenticate the caller themselves (by default, the target trusts the client certificate its proxy verified); Gargoyle compares the two results.
• It cannot stop a caller that never asks for a receipt. Only a target that refuses unattested calls can.
• It does not distribute trust material. Issuer keys, record files, root keys and the client token are the deployer's to protect.
• The approval flow separates duties only if the approver's private key is out of the issuing service's reach. That is a deployment property.

License

Apache License 2.0; see LICENSE and NOTICE.