FlurryPORT

Set up or rotate signed-intent posting for an endpoint. Inputs: an optional header the signature rides in, and projectId plus endpointId. A high-entropy signing key is generated LOCALLY, stored in the CLI keystore, and registered with the server in one atomic step, so the key value never appears in this conversation. Afterwards post_intent signs automatically and the endpoint rejects an unsigned post with 401 before storage. Re-running ROTATES the key, and anything still signing with the old one breaks at that moment. Read-write token.

Set the room map. Marks captureId as the endpoint's current ORIENTATION, the one retention-exempt capture every room keeps on every plan, and stores sections and roster beside it. Owner only, read-write token. A new orientation releases the previous one, and seats read it from get_endpoint OrientationCaptureId then get_capture. sections and roster come back from list_sections, get_canon, and get_endpoint, so a reader never parses the orientation post. Pass clear true with no captureId to release the orientation, or sections or roster alone to edit the map without moving it. Omitting one leaves it as it is; an empty array clears it.

Set the COMPLETE custom header set on a replay target. Full replace: include every header you want kept, and an empty array clears them all. Inputs: targetId, headers in send order, and projectId plus endpointId. This is how a recipe targetTemplate.headers get wired. Delivery applies them on top of the captured request headers, overriding case-insensitively. CREDENTIALS NEVER GO IN LITERALLY: write the vault reference $secrets.NAME for the secret part. Scheme words stay literal, so "Bearer $secrets.SLACK_BOT_TOKEN" is right and a pasted token is refused server-side. Then call request_secret_setup so the user supplies the value. Benign literals such as Content-Type and version pins are fine. Verify with get_replay_target deliveryHeaders. Read-write token.

Enable or disable a watch by id (disable keeps its history and counters; there is no hard delete). Get ids from list_watches.

Start or reuse a local echo receiver on the user's machine, so replay can be proven before their real backend exists. It answers 200 and mirrors the method, headers, and body back, so a forward_to_localhost against it SHOWS the delivered webhook in responseBodyPreview. Input: an optional port, default 4242, falling back to an ephemeral one if busy. Runs inside this MCP process on loopback only and stops when the session ends. Repeat calls return the running instance.

Update a replay target in place, most importantly ARMING it with autoReplay true so standing bindings and fan-out deliver through it, or disarming it. Inputs: targetId, autoReplay, name, baseUrl, and projectId plus endpointId. Any omitted field is left unchanged. Use it instead of creating another target: bindings stay attached and plan limits count targets. Plans cap how many targets may be ARMED at once, so disarm one before arming another if the server refuses. Read-write token.

Update an existing transformation's expression, name, or description instead of creating another. Plans cap transformations per endpoint, so revising the one you have is the right move when a draft was wrong. Inputs: transformationId, an optional name, description, and expression, and projectId. Versioning is copy-on-write: when the latest version is frozen by a standing binding, the update FORKS a new draft, so check forkedNewVersion and re-bind with the returned latestVersionId. Materialize $install.NAME placeholders before saving; binding refuses an unmaterialized template. Read-write token.

Verify a hash-chained event log in ONE call, so you never hand-roll canonical JSON or a SHA-256 chain. Inputs: events, the parsed bodies in sequence order oldest first, and optionally hashField and genesisPrevHash. Returns intact, brokenAt for the index of the first break, count, and nextPrevHash, the exact prevHash for YOUR next event. Canonicalization is RFC 8785 JCS: keys sorted at every level, top-level sig excluded, nested sigs hashed. Pure computation; needs no account.

Block until a new capture lands on the endpoint or the timeout elapses, then return what is new since your cursor with bodies inline, exactly like list_captures with after. Use it for a DEDICATED monitor whose only job is to follow a stream; to check between other work, prefer list_captures with after, which returns immediately. Inputs: after, a NextCursor from any read, timeoutSeconds up to 25, includeBody, and endpointId. Returns immediately if captures already landed since the cursor, so you never miss events between calls.

Record or remove a pipe entry in .flurryport/pipes.json, upserting by name, so the wiring survives restarts and travels with the repo. NEVER put a key or token value in any field: a signing key is referenced by localKeyRef only, and the write is refused if anything token-shaped appears. Mark draft true while wiring, and clear it once the pipe is verified.