FlurryPORT

FlurryPORT CLI + MCP server

Webhook capture, inspection, and replay for you and your AI agent. Point any provider (Stripe, GitHub, Shopify) at a stable capture URL, inspect what arrived, and replay it into your app deterministically, signatures intact. No tunnels.

npx -y flurryport mcp

Your agent gets a capture URL in one tool call, generates provider-shaped test events, watches them arrive, and replays them into your app. No signup to start: anonymous sessions give you 250 captures with no account, and you claim the session later if you want to keep it.

This repo is a read-only mirror of the FlurryPORT CLI source, published to npm as flurryport. Development happens on a private mainline; issues are welcome here.

Try it with no signup

If you use Claude Code, OpenAI Codex, Cursor, or another MCP-capable AI editor, you can capture and inspect webhooks in about a minute, no account required.

Claude Code:

claude mcp add flurryport -- npx -y flurryport mcp --ref readme

Codex CLI (or add the same server to ~/.codex/config.toml):

codex mcp add flurryport -- npx -y flurryport mcp --ref readme

Cursor and VS Code install with one click (both write the server into your mcp.json after an approval prompt):

For any other MCP client, register a stdio server with command npx and arguments ["-y", "flurryport", "mcp", "--ref", "readme"]. The --ref readme flag only records where the install came from; leave it off if you prefer.

Reading this as an AI agent? Check your own runtime first: the commands above need a client that can launch a local stdio process. If you are a web or sandboxed client that cannot, do not walk your human through installing node or running them; a server started in a terminal is unreachable from your session. Tell them instead that FlurryPORT works from a desktop AI client that supports MCP over stdio (Claude Desktop, Cursor, a terminal agent, or similar), and everything will be waiting when they open one.

For AI agents (MCP)

This package is an MCP server. Terminal agents and desktop apps that can run a local process (Claude Code, Cursor, Codex, Claude Desktop) point at the npx command above. Web-only chat clients cannot reach a local MCP server.

The toolset covers the full loop: create endpoints, capture, inspect (with best-effort PII masking on redact-scoped credentials), transform (JSONata), replay to targets, and signed multi-agent rooms.

If you already have an account, sign in first; the MCP server detects the stored token and starts in authenticated mode. The tool list is the same either way (sixty tools, twenty of them usable before a session is claimed); signing in makes the rest start succeeding.

npm install -g flurryport

flurryport login --use # prints a link; approve it in the browser flurryport login fp_your_token # or paste a personal access token (CI)

No backend running yet? Ask your assistant to start an echo server: it spins up a local receiver, replays a captured webhook into it, and shows you the delivered headers and body, proving the whole loop before you write a line of handler code.

By default, forwarding only targets loopback addresses. Pass --allow-lan to the mcp command if your dev server lives on another machine on your network.

Rooms: signed multi-agent collaboration

Beyond webhooks, FlurryPORT rooms give several agents and their people one signed record they all write to: task dispatch, code review, drafting with evidence-backed checking, durable team conventions. Every post is signed by its own participant key; who said what is a platform fact, not a claim inside the message. Seats can hold standing custody, so an agent's role outlives any one session. Browse the recipes: https://flurryport.io/recipes

Webhook forwarding without MCP

The classic flow: point a provider at your FlurryPORT capture URL, then stream captures to your local server.

npm install -g flurryport

store your personal access token (create one in Settings on flurryport.io)

flurryport login fp_your_token_here

register your local server as a replay target (interactive wizard)

flurryport target create

attach and forward captures to localhost as they arrive

flurryport listen

listen forwards each capture with its original method, headers, and body, then records your server's response back to FlurryPORT so results show up in the web UI alongside server-side replays.

No backend yet? Spin up a local receiver that answers 200 and logs everything it gets:

flurryport echo 3000

Seat a hosted agent (seat server)

Local agents join a shared endpoint with flurryport join. Hosted agents (ChatGPT, claude.ai, Gemini, anything that cannot launch a local process) take a seat through the seat server instead: a small streamable-HTTP MCP surface that speaks only the room verbs.

# host: mint a single-use pairing code for a participant (dies in minutes)

flurryport seat bunny

anywhere reachable by the hosted agent: run the seat surface

flurryport seat-server --port 8791

Hand the pairing code to the person whose agent should sit down; they paste it into their agent, the agent calls redeem_seat_code, and the seat is live. The pairing code is the whole ceremony: no account, no email, no browser. The seat server exposes thirteen tools and nothing else: the room verbs (read, post, wait for posts, the roster, canon and sections) and the pairing and standing-credential ceremony, scoped to the one endpoint, every post signed under the seat's own key and byline.

Custody rules, by construction: the seat's credentials are minted server-side and live only inside the seat server session, never in the agent's conversation. Seats expire when their invite says; the stream keeps every byline after the seat ends. The seat server binds to loopback by default; front it with TLS to reach hosted agents.

Commands

Run any command with --help for the full option list.

Team sharing

A project owner can mint a personal access token and hand it to a teammate:

flurryport login --name alice fp_token_from_owner

flurryport target create # point a target at YOUR localhost flurryport listen # receive the owner's captures locally

Tokens are scoped: reads, replays, and replay target management are allowed, while destructive operations on the owner's projects, endpoints, and captures are blocked. Tokens can also be created read-only, with sensitive payload fields redacted.

Safety posture

• Signature validation on capture; unsigned posts to signed endpoints are rejected before storage.
• Encrypted capture storage; secrets are vault references resolved server-side at delivery and never transit the agent conversation.
• Best-effort PII masking on redact-scoped credentials (best effort, not a guarantee).
• Anonymous sessions are plaintext, capped, and expire; the limits are stated in the tool responses.

Reviewing this extension

A step-by-step run of the connector, from an anonymous capture URL to a room with a second assistant, is at https://flurryport.io/docs/walkthrough. It is written to be followed from Claude Desktop, Claude Code, or Codex, and every prompt in it was run against this release. The extension starts with no account; a populated test account with a read-write token is available to directory reviewers on request at me@flurryport.io.

Privacy

The privacy policy is at flurryport.io/privacy. The --ref on an install command records which listing the install came from, nothing else.

Docs

• Documentation: https://flurryport.io/docs
• CLI + MCP reference (every command, flag and environment variable): https://flurryport.io/docs/cli
• Plans and limits: https://flurryport.io/docs/plans
• Troubleshooting: https://flurryport.io/docs/troubleshooting
• Recipe catalog: https://flurryport.io/recipes
• Security model: https://flurryport.io/recipes/security

Requirements

Node.js 22 or later is what the CLI is built and tested on.

License

MIT