The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Agent Mandate Verification listing page.
Check a signed authorization envelope before trusting an agent's claimed authority.
Agent Mandate's MCP tool accepts the complete { mandate, signature } envelope
returned by POST /v1/mandates. Raw claims are not a signed envelope.
The walkthrough: create an envelope, verify an action through MCP, then change one signed field and verify again. You see the real verification result for each request, including the one that gets refused.
Requirements: Node.js 20 or newer, and an Agent Mandate API key.
Getting a key: the free tier includes 500 verified actions per month and needs no card. Issuing mandates is free and never consumes the allowance; one unit is one action verified. Paid plans start at $299/month for 10,000 verified actions. This walkthrough consumes at most two of your free units.
Running this in production? When a signed mandate is and is not worth it, with the published pricing. Short version: if the approval gate is one you control, a policy check inside your own service is simpler and cheaper. The signature earns its price when somebody other than you has to be able to verify the decision.
Just exploring? POST /v1/demo/verify takes raw claims, needs no key at all, and
is shown at the end. It does not replace the signed-envelope walkthrough, because it
does not check a signature.
The example runner comes from this repository, but it installs
@api-disk-integrations/agent-mandate-mcp@0.1.1 from npm into a throwaway prefix
and runs that, so it exercises the published artifact rather than your working tree.
Omit AGENT_MANDATE_API_KEY and it prompts without echoing.
That is a real run against production, not illustrative output. The grant allows up to
100,000 minor units but requires approval above 25,000, and the action asks for 30,000,
so requires_approval is the correct answer. A deny or requires_approval is a
correct result, not a failure.
Verification reports a result; your application remains responsible for enforcing it.
A generic stdio client configuration:
${AGENT_MANDATE_API_KEY} denotes the host's secret reference; use your client's
documented secret facility if its syntax differs. The package uses stdio and reads
exactly that environment variable. It has no remote /mcp endpoint.
Pin the version. 0.1.0 is still on the registry and has a first-use defect: its
input schema accepted any object for mandate, so a call built from the keyless
demo's shape returned HTTP 400.
If you would rather verify a checksummed artifact, every release also attaches a tarball and its SHA-256:
| Endpoint | Key | Takes |
|---|---|---|
POST /v1/demo/verify | none | {mandate: {…claims…}, action: {…}} — raw claims |
POST /v1/verify | yes | {mandate: {mandate: {…claims…}, signature: "…"}, action: {…}} — the envelope |
The envelope is the entire response body of POST /v1/mandates:
{mandate, signature, requestId}. Pass it through unchanged.
"mandate.mandate" must be the claims object means bare claims were passed where the
envelope belongs. Run POST /v1/mandates first and pass its whole response.
Step 1 — create a mandate. Issuing is free and does not consume your allowance.
Answers 200 with {"mandate": {…}, "signature": "…", "requestId": "…"}.
That whole body is the envelope.
Step 2 — verify an action against it. Call verify_action with the envelope as
mandate:
This route takes bare claims, not the envelope, and does not check a signature.
It does not issue or revoke mandates, execute an action, change an account, or call billing. It makes at most one API request per tool call and never retries automatically.
@api-disk-integrations/agent-mandate-mcpio.github.API-Disk-Integrations/agent-mandateAPI-Disk-Integrations/agent-mandate-mcpThere is exactly one tool, verify_action. Supply either one action or an
actions array, never both. A batch contains 1–500 actions. Monetary amounts
are non-negative integer minor units and require currency.
Example input:
Representative successful structured output:
Treat all three decisions literally. In particular, neither allow nor
requires_approval executes the proposed action.
AGENT_MANDATE_API_KEY is absent, the tool returns an MCP error before
making a request.https://agentmandate-api.com/v1/verify endpoint. There is no automatic
retry; Fetch uses explicit redirect: error, and redirect responses are
rejected without a follow-up request.Content-Length representation.2026-07-28 and retains the SDK's listed 2025 compatibility
revisions.The source link is not evidence that the npm package or Registry listing is available. Those two releases require their own public readback. A clone, install, download, tool call, or listing is not evidence of customer activation or revenue.