Skip to main content
Glama

drift

M8ven Score Listed on mcpservers.org

Continuous mandate enforcement for AI agents. Every action is checked against the signed mandate before it happens, not once at signing.

pip install drift          # or: pip install -e .
python3 -m drift.http_server --http 8080

The gap

An agent is granted a mandate: who to pay, how much, in which currency, until when. AP2 verifies that mandate once, at signature time. After that the agent acts alone, and nothing holds it inside its authority.

Measured proof that this matters, from the AP2 Python SDK itself:

verify_chain documents:
  "enforcing expected_aud / expected_nonce on the terminal hop when provided"

but kb_sd_jwt.verify calls verify_expected_claims only when typ in TYP_TERMINAL,
and a kb+sd-jwt+kb token — which the SDK itself produces whenever the payload
carries a cnf — falls into TYP_INTERMEDIATE:

  hop type      : kb+sd-jwt+kb
  aud in token  : https://attacker.example/ap2
  expected_aud  : https://honest-merchant.example/ap2
  ACCEPTED a token with the wrong audience.

The protocol an agent trusts with its spending authority does not check that the authority was not substituted. drift checks it.

Related MCP server: Agent Receipts

What it checks

#

rule

catches

1

amount

over the limit — the most common form of drift

2

payee

wrong recipient

3

audience

payment redirected to someone else

4

currency

currency substituted

5

window

outside the validity period

6

intent

intent outside the allowlist

Default is DENY. Rules run in fixed order, so a refusal names the most important reason rather than an incidental one.

MCP tools

tool

purpose

drift_check

check one action against a mandate

drift_session

check a sequence, get a report with a chain fingerprint

drift_explain

show the rules in the order they apply

No SSE, no stateful sessions, no HTTP framework. Standard library only.

Example

from drift import Mandate, Drift, Action

m = Mandate(
    payee_id="shop", audience="https://merchant.example/ap2",
    max_amount_minor=10_000, currency="USD",
    valid_from=1_700_000_000, valid_until=1_700_086_400,
    allowed_intents=("buy",),
)
d = Drift(m)
verdict, rule = d.submit(Action(
    intent="buy", payee_id="shop", audience="https://merchant.example/ap2",
    amount_minor=99_999, currency="USD", at=1_700_000_060,
))
print(verdict, rule)   # DENY amount

Limits, stated plainly

  • Does not verify mandate signatures. That is AP2's job. drift runs after signature verification, at the level of decisions.

  • Does not execute. It decides; the caller acts.

  • Does not keep state. A session is an argument. No state means no attack surface and no divergence under replication.

  • Does not tell truth from intent. intent arrives from the agent and is untrusted input. drift bounds the damage; it does not certify the intent.

If an agent controls both intent and amount, it can lie in both. Closing that gap is attest's job: attest proves afterwards, drift bounds beforehand. They are complements.

Use it from your agent

{
  "mcpServers": {
    "drift": { "url": "https://drift.rta.lat/mcp" }
  }
}

That block is how our own agents call this server. Tools: drift_check (one action), drift_session (a sequence), drift_explain (the six rules and why each exists). 100 free checks per day, then $0.005 per call over x402.

Licence

Apache-2.0.

Related MCP Connectors

Related MCP Servers

  • A
    license
    A
    quality
    C
    maintenance
    Cryptographic proof of consent for AI agents. Sign before you act. Policy engine enforces spending caps, action whitelists, and escalation rules. Independently verifiable by anyone.
    10
    2
    Apache 2.0
  • A
    license
    A
    quality
    D
    maintenance
    Enables AI agents to make payments with strict spending limits enforced by signed mandates, preventing unauthorized fund movement beyond predefined caps and allowlists.
    4
    12 npm
    MIT
  • A
    license
    Not graded
    quality
    D
    maintenance
    Provides cryptographic governance receipts for AI agents, enabling pre-execution evaluation and signed verdicts (EXECUTE/BLOCK/REVIEW/SHADOW) with offline-verifiable audit trails.
    MIT