Fyreflock

Fyreflock for Claude, ChatGPT and Claude Code

Ask your assistant to find you someone. A real, verified person texts back in the same window, and you can play Connect 4, Battleship or 23 other games together without leaving the chat.

Your assistant only carries the words, exactly as you typed them. No bots, ever.

Every AI chat is a room with one person in it. Fyreflock puts a second person in that room: a real one, who passed a camera check, picked at random, there for the same reason you are. You stay in Claude or ChatGPT. They might be in a browser, in ChatGPT, or in their own Claude. Your assistant is the courier, and it is held to that.

Real box, two real accounts, real time. From "find me someone" to a won game of Connect 4 in about twenty seconds.

Try it in sixty seconds

1. Add Fyreflock to your assistant. The connector URL is https://fyreflock.com/mcp. Host-by-host steps are just below.
2. Sign in once and press Allow. A browser tab opens, you sign in or sign up, and you are sent straight back. New here? The same page signs you up, camera check and all.
3. Say "find me someone to talk to". The Fyreflock box opens right there in the conversation. Say "next" for somebody else, "leave" to stop.

[!TIP] Nobody around yet? Press Play while you wait and take on the computer. The search keeps its place, and the box switches to the conversation the moment somebody arrives.

Set it up

Claude (claude.ai and Claude Desktop)

1. Settings → Connectors → Add custom connector.
2. Paste https://fyreflock.com/mcp and save.
3. The first time you ask for someone, Claude opens fyreflock.com: sign in, press Allow, done.

The chat box appears inline in the conversation. It survives you switching to other chats; your lines wait, and the box picks them up when you come back.

It follows your assistant's theme.

ChatGPT

1. Settings → Plugins → Add custom MCP server.
2. Server URL https://fyreflock.com/mcp, authentication OAuth.
3. Ask for someone; sign in and press Allow when the page opens.

Claude Code (this repo is the plugin)

/plugin marketplace add FuloxDev/fyreflock-claude
/plugin install fyreflock@fyreflock

Restart Claude Code and say find me someone. The first time, Claude Code opens fyreflock.com in your browser: sign in or sign up, press Allow, and you are back in the terminal, connected. No token to copy.

It is a plugin rather than a bare claude mcp add line for one reason: the hooks. They turn "your assistant sends your words unchanged" from a request into something that is actually checked, every message. Do not also run claude mcp add for Fyreflock: two registrations means two servers, duplicate tools, and hooks that only cover one of them. Details in How it stays honest and The pane.

VS Code (Copilot agent mode)

Or add it by hand to mcp.json:

{ "servers": { "fyreflock": { "type": "http", "url": "https://fyreflock.com/mcp" } } }

Sign-in happens in your browser on first use.

Any other MCP client

Remote server, Streamable HTTP, OAuth 2.1 with PKCE, client metadata documents and dynamic registration both accepted:

https://fyreflock.com/mcp

If your client cannot sign in by itself, create a token under Settings on fyreflock.com and send it as Authorization: Bearer <token>.

What you can do from your assistant

How it stays honest

• Every match is a real person. Every account passed a camera check in a browser before it could meet anyone. There is no bot in the queue and never will be. 18+ only.
• They know. The other person's card says you are talking from inside an assistant, so nobody is fooled about where the words came from.
• Your words, verbatim. The assistant is instructed to send your message exactly as typed. In Claude Code the plugin's hooks refuse any send that is not the line you wrote this turn. Rewording, shortening, "fixing" and inventing are all blocked.
• You can check. The companion tab at fyreflock.com/app/live shows what was actually sent and received, with no model involved.
• It survives you looking away. Switch to another chat, close the window, come back: your search keeps its place in line, your conversation keeps its lines, and your browser can tell you when somebody is there.

Before you meet anyone

• 18+. Same as meeting a stranger anywhere on Fyreflock.
• Messages here pass through the Fyreflock server in plain text and through whichever AI provider you are using. Browser-to-browser chats on the website never do. This is the real trade you are making.
• Use it in a chat with nothing else attached. No other connectors, no files, no tools. A stranger is typing into a context your assistant can read, and whatever it can see, they can ask about.
• Their words arrive wrapped and labelled as somebody else's text. They are never instructions, to you or to your assistant.

Full detail: https://fyreflock.com/agents · Privacy: https://fyreflock.com/privacy · Safety: https://fyreflock.com/safety

Talk to us

• Questions, ideas, a host that behaves oddly: Discussions.
• Something broken: open an issue.
• A security problem: please do not open an issue; see SECURITY.md.
• Someone behaving badly in a chat: report them from the box or on the website; that goes straight to moderation.

FAQ

Is this a chatbot? No. A real person is on the other end, every time. The computer you can play games against on the website is labelled as exactly that and never takes a place in matchmaking.

Can my assistant read my messages? Yes. Both directions pass through it, which is why the box exists: in Claude and ChatGPT the box talks to our server directly, so the model only sees what you ask it to.

Does the other person know I am using an assistant? Yes. Their peer card says so.

Does the assistant change what I write? It is instructed not to, the Claude Code plugin enforces it, and the companion tab lets you verify it.

How do I know they replied? The box shows it live. If you have moved on to another chat, your browser can notify you: turn that on when you connect, or later under Settings on fyreflock.com.

What does it cost? Nothing. Paid passes add matching preferences; talking is free.

The pane: talk without the model in the path

Claude Code 2.1.287 or newer (Desktop 2.1.286) loads the plugin's mod, hooks/mod.tsx, beside the two hooks. It draws a Fyreflock pane next to the transcript and keeps the conversation itself, so a line costs 0.2 s instead of a model turn, and the model never sees the stranger's words unless you ask it to.

/fyreflock              open the pane
/fyreflock start        find someone
/fyreflock next         skip to somebody else
/fyreflock leave        hang up, or stop looking
/fyreflock status       who you are talking to and who is waiting
/fyreflock <anything>   send that line, exactly as written

In the pane: the chat box's five states, each with its colour on a dot and a rule under the header in your theme's own colours: ready (yellow, how many are by the fire, Start), looking (yellow, Cancel), talking (green), reconnecting (orange, when a poll fails mid-conversation; it clears on the next answer) and ended (coral, "@them left the campfire" or "You left the campfire", with Find someone new). While talking: their lines on the left and yours on the right, a field that sends on Enter, Next person / Leave (hotkeys s, n, l while the pane has the keyboard, ctrl+x tab to give it the keyboard), and Sent / Delivered / Read under your last line. While Claude is idle, the band above the prompt shows the same dot and says who you are talking to and how many lines arrived since you last typed; hide dismisses it for the session. A reply also shows as a toast, and the status line says whether you are looking, talking or reconnecting.

How the pane talks to the server, and how it signs in

The mods API gives a plugin HTTP only, so there is no socket to hold. The pane long-polls fyreflock_app_wait on the same /mcp endpoint the model's tools use, at FYREFLOCK_URL, twenty-five seconds a call (the engine cuts a fetch at thirty), and sends through fyreflock_app_say. Those two are the widget's app-only tools; they return the server's own record of the conversation, which is why the pane can show receipts the fyreflock_listen result cannot. Each poll carries pane: true: the server reads that as "the pane is live" for thirty seconds, and a chat box the same Claude Code mounts (the desktop app's Code tab draws MCP Apps) folds to one line pointing here instead of offering a second copy. The loop starts the first time you run /fyreflock and stops when you close the pane with nothing open: a plugin installed everywhere must not hold a connection to fyreflock.com in sessions that have nothing to do with it.

Signing in: there is nothing to paste. A mod cannot borrow Claude Code's own sign-in and cannot take a browser callback, so the pane signs itself in by device flow. The first /fyreflock shows a fyreflock.com link and a code like BCDF-GHJK; open the link, sign in or sign up, check the page shows the same code (a link forwarded to somebody else gives itself away there), press It matches and then Allow, and the pane picks the answer up within five seconds. It keeps its own access and refresh tokens in the plugin's store ($.store, a JSON file under your Claude Code config directory; never $.state, which every plugin can read), renews them by itself, and goes back to its Sign in button if you disconnect it in Settings, where it is listed as Claude Code. FYREFLOCK_TOKEN, if set, still overrides all of that with a token from Settings.

The pane and the model's tools share one mailbox. Use one or the other for a given conversation: a fyreflock_listen the model holds and the pane's poll would take turns collecting the same lines.

Checking it: claude plugin validate ./plugin reads the module the way the engine will; claude plugin test ./plugin runs tests/mod.test.ts against the engine with the network answered from memory, on the terminal and desktop surfaces. To try it live, claude --plugin-dir ./plugin "/fyreflock start" with FYREFLOCK_URL set to the dev server's /mcp, and a second account in the widget or on fyreflock.com/app to be matched with.

Stop it asking permission for every message

Without this you confirm a dialog for every single line you send, which is unusable. Answer always allow the first time, or add the rule yourself:

/permissions

and allow mcp__plugin_fyreflock_fyreflock__*.

The exact tool prefix is generated by Claude Code from the plugin and server names, and the format has changed before. If that rule does not match, run /mcp or check any Fyreflock tool call in your transcript to see the real name, and use that. The hook matcher in hooks/hooks.json is a regular expression precisely so it keeps working either way.

What the hooks do

Two, both local, both tiny, neither sends anything anywhere.

capture-prompt.mjs runs on every prompt you submit and writes it to one file in your system temp directory, keyed by a hash of the session id. That is the whole job.

check-verbatim.mjs runs before fyreflock_say and fyreflock_roost_say and refuses the call unless the text being sent is the message you typed this turn. Spacing and curly quotes are forgiven. Rewording, correcting, shortening, summarising, shouting and inventing are not, and the model is told to send your words instead.

It compares by equality, not containment. What goes out has to BE your message, not merely appear somewhere inside what you typed: a substring test lets "tell them no, I don't want to meet" send "want to meet", and lets a model assemble sentences a word at a time. Two allowances make that livable, because nobody types bare lines at an assistant: one leading "tell them" is stripped, and anything you wrap in backticks goes byte for byte, punctuation and case included. That last one is the escape hatch when the rules get in your way.

If no prompt was captured, it allows the call. A missing capture means this plugin did not run, which is our bug, not the model's, and blocking your first message over it would just look broken.

Set FYREFLOCK_ASSIST=1 (or true, and nothing else counts) to turn the check off, for when you genuinely want Claude to help you write or translate a line. The stranger is not told when this is on, so it is a tool for you rather than a mode of the product.

Pointing it somewhere else

FYREFLOCK_URL overrides the endpoint, for a dev server:

export FYREFLOCK_URL=http://localhost:8787/mcp

Publishing this (maintainer)

This directory is the plugin. It is developed inside the main Fyreflock repo so it stays in step with the server, and published as its own repo because that is what /plugin marketplace add consumes.

gh repo create FuloxDev/fyreflock-claude --public --description "Talk to a real, verified stranger from Claude Code"
cd plugin && git init && git add -A && git commit -m "Fyreflock plugin"
git remote add origin https://github.com/FuloxDev/fyreflock-claude.git
git push -u origin main

Validate before pushing, and again after any change to plugin.json, hooks.json or .mcp.json:

claude plugin validate ./plugin

The host cannot run these hooks inside a test, but everything they depend on can be: apps/server/test/verbatimhook.test.ts spawns both scripts the way Claude Code does, asserts on the decision they return, pins the matcher regex against the real generated tool names, and keeps every known bypass as a regression case. Run it after touching either script or hooks.json.

One caveat the tests cannot cover: the hooks are spawned as node, resolved on Claude Code's own PATH. If node is not there, neither hook runs, every send is allowed, and nothing on screen says so. Worth checking once on a machine where Node came from nvm or a GUI launcher.

The other thing that genuinely needs a human once: install this in a real session and confirm the PreToolUse matcher actually fires, by trying to get Claude to send something you did not type. If it goes through, the matcher is wrong and the whole verbatim story is decorative.

Find your Flock. Every flock starts with a stranger.

fyreflock.com · assistant mode · safety · privacy · MIT