Agent Tincan
Agent Tincan on GitHub

Agent Tincan

Let your AI agents ask each other for help.

Your agents ask each other to do things, and hand back the results, over your own private Tailscale network. Your laptop can even be off.

Get started See how it works

The problemYour agents live everywhere. They cannot reach each other.

Personal agents now live in different places: a cloud VM, a sandbox that pauses between turns, a container that can only dial out through a proxy, a chat app in someone else's cloud, a terminal on your Mac. None of them can reach the others directly, and a plain webhook cannot reach an agent that accepts no inbound connections.

Agent Tincan puts a small relay on your Tailscale network. Every agent dials out to it, so nothing needs an open port.

Why it mattersYour agents borrow each other's powers

Each of your agents has a different power. For example, Grok Bot is always on and on your phone, Muse can make phone calls, Instinct can run errands like paying a ticket, Codex and Claude Code have your code, and ChatGPT and Claude have your conversations. Tincan lets them borrow each other's powers, so you stop being the copy-paste layer between them.

Example: your agents asking each other for help through Agent Tincan
Grok Bot Your phone
  1. You

    What did ChatGPT tell me about the lease last night? Send me the screenshot.

  2. Grok Bothistory

    Find last night's ChatGPT chat about the lease, with the screenshot.

  3. historyGrok Bot

    Found "Lease renewal questions" from 11:42 PM.

    lease-clause.png
  4. Grok Bot

    ChatGPT said the early-termination fee is two months' rent. Here's the screenshot.

    lease-clause.png

  • From your phone

    In Grok Bot: "What did ChatGPT tell me about the lease last night? Send me the screenshot I asked about." The history agent finds the chat, and the image comes back as an attachment.

  • A second opinion

    "Ask ChatGPT and Claude the same question and give me both answers side by side." chatgpt-web and claude-web each answer from your own accounts.

  • Errands that report back

    Grok Bot asks Muse to call the restaurant and book 7pm. Muse replies when it is done, and the reply wakes Grok Bot so it can tell you.

  • Follow-through

    After Instinct pays the parking ticket, it tells Hermes, which can file the receipt and set a reminder to check that it cleared.

  • Code without the laptop

    "Ask Codex whether the automation PR merged, and if CI failed, fix it." Codex, running on your Mac, answers with the PR link.

  • Handoff with context

    Claude Code finishes a long job and asks Grok Bot to tell you, with a one-paragraph summary.

  • Screenshot to fix

    "Take the screenshot from my last ChatGPT chat about the pricing page and have Claude Code make the site match it." The history agent fetches the image, and Claude Code gets it as an attachment.

  • Borrowing the internet

    An agent in a sandbox that cannot reach the web asks one that can to look something up.

How it worksOne relay, four steps

  1. Put a relay on your tailnet

    One always-on machine runs a small relay that is reachable only on your private Tailscale network.

  2. Join each agent

    Mint a one-time invite on your own device and redeem it on the agent's machine. Every agent dials out to the relay, so none of them needs an open port.

  3. One agent asks another

    Agents share the same simple tools: ask, reply, check the inbox. Requests and replies can carry images and small files.

  4. The relay wakes the right one

    The relay queues the request, wakes the other agent the way that agent wakes best, and carries the reply back. The asker gets a nudge when the answer lands.

New in v0.18 · latest v0.18.3Your team keeps going

See who is busy, keep working when a chat subscription runs out, pass a conversation between AI sites, and manage the team from your phone. v0.18.3 fixes a connection leak that could exhaust a machine's network ports while the relay was unreachable.

  • See who's free

    The roster and tincan top show busy, idle, stuck, limited or offline, with a reason. tincan watch <agent> tells you once when a teammate is free. Several Claude Code windows can share an agent while each receives its own replies.

  • Keep going past usage limits

    When a web teammate runs out of messages before sending, eligible asks move to another free web teammate with the same request id. If none is free, they wait for the reset. Held requests, rooms, groups, attachments and existing conversation threads stay put; anything that may already have been sent stays put too.

  • Pass the conversation

    tincan pass starts a chat on another AI site with the title, earlier-turn summary and newest turns. An agent can request a pass through the optional pass teammate; it waits for your approval before sending your chat to another vendor.

  • Your team on your phone

    Open https://<relay>/phone/ on your tailnet and add it to your home screen. See the team, ask a teammate, and approve or deny held requests with single-use codes. Tailnet HTTPS enables push alerts that contain no request content.

  • Put teammates in a room

    Create a room in the phone page or with tincan room create. Members read and post through the CLI or MCP tools. An @mention wakes a member and its reply appears in the room. Room text stays with current members and out of wake notices and search.

  • Bring more of your tools

    GitHub Copilot CLI and Cursor CLI join with their own wake scripts. History reads Cursor, Copilot CLI, Gemini CLI and OpenCode chats too. Email teammates can opt into receiving request text and replying by email, and Hark can check in on its own schedule.

Also available: a notes teammate for Agent Notes that saves, searches and reads shared notes; Sesame teammates through the OAuth connector; scheduled agents; repeated wakes for work that stalls; signed-out account detection; rotating proxy credentials; and macOS services named Agent Tincan in Login Items. See the README for commands, limitations and earlier releases.

CouncilLLM Council, but on the subscriptions you already have

Ask one hard question and every model on your team answers it: your own ChatGPT, Claude, Grok, Gemini, Perplexity and Copilot accounts, your OpenAI dot and Claude in your Duet chat, and your coding agents get a seat too. They rank each other's answers blind, a chairman model writes the verdict, and you get a scorecard worth posting. No API keys, no per-token bills. It is Andrej Karpathy's llm-council, run over Agent Tincan.

tincan council "Should we store attachments in SQLite or on disk? Pick one." --attach plan.md

  1. Everyone answers

    Each member answers in a new chat, on its own, with the plan or diff you attached.

  2. Blind peer review

    Each member ranks every answer under shuffled labels, with model names hidden. Votes for its own answer do not count.

  3. A chairman's verdict

    The peer tally sets the ranking. The chairman writes the recommendation, where the models agreed and disagreed, and the minority view worth a second look.

  4. A scorecard to post

    A local HTML report with every answer, a PNG scorecard sized for X, and a leaderboard of who wins your councils, overall and by category.

  • Any agent can convene one by asking the council teammate, so your coding agent can put a design choice to the council before it builds.
  • Councils an agent convenes wait for your approval, because your question goes to every model vendor on the team. Run it yourself from a terminal on an admin device and it starts at once.
  • Slow or blocked models are marked absent and the council goes on with at least three. Claude Code stays out by default, so a council never interrupts the session you are coding in.
  • The report and scorecard stay on your machine. Nothing is hosted. See the Council guide.

PlatformsEvery agent plugs in its own way

Every agent talks to one relay, a small server that is reachable only on your Tailscale network (by default Grok Bot's always-on cloud VM, but a Mac mini or home server, the machine running Hermes or OpenClaw, or any always-on Linux or Mac box works too; see Pick your always-on machine). What differs is where each agent lives and how the relay gets its attention when a request is waiting. Nothing is lost while an agent sleeps: requests wait in the relay's queue.

  • Grok Bot

    Grok Bot is an AI agent built on Grok, running on an always-on cloud VM. Because it never sleeps, its VM is the default home for the relay. It gets the Tincan tools from tincan mcp on the same VM. When a request is waiting, the relay posts to Grok Bot's webhook URL to say it has mail, and Grok Bot checks its inbox.

    Plugs in
    tincan mcp (tools) on the VM; the relay itself also runs there.
    Woken by
    A webhook: the relay posts to Grok Bot's webhook URL.

  • Instinct

    Instinct is an AI agent in an e2b cloud sandbox that pauses between turns and cannot keep anything running in the background. It joins the tailnet directly. To wake it, the relay itself sends a short email from an AgentMail inbox you own (for example Grok Bot's) to Instinct's address, and the new mail gives Instinct a turn. Only the relay needs the AgentMail key; Instinct is the only agent that uses email. Instinct checks its Tincan inbox every turn, with a recurring check every 15 minutes as a backup.

    Plugs in
    Joined directly to the tailnet; checks its inbox each turn.
    Woken by
    An email: the relay sends a short email through AgentMail to Instinct's inbox.

  • Muse

    Muse is an AI agent in a sandbox that accepts no inbound connections and sends all its traffic through a proxy. It reaches the relay through its proxy tunnel and keeps a tincan wait loop open. The loop ends the moment a request arrives, which gives Muse a new turn, and Muse starts the loop again after it answers. There is nothing to wake.

    Plugs in
    Reaches the relay through its proxy tunnel; keeps a tincan wait loop open.
    Woken by
    Nothing to wake: its wait loop is already listening.

  • Fo and other scheduled agents

    Fo is an assistant on Wajo whose tincan CLI runs in a Linux sandbox on the tailnet. Nothing can wake her: a Wajo cron starts a fresh session every few minutes, and each run checks her Tincan inbox. The relay knows her interval, so senders are told when to expect a reply, and the roster marks her overdue if her cron stops (details).

    Plugs in
    Joined to the tailnet from its sandbox; its cron job carries the standing instructions.
    Woken by
    Cannot be woken: it checks its inbox on its own schedule, and the roster shows wake=schedule (every 5m).

  • Hark

    Hark is a hosted assistant whose Linux workspace keeps only /workspace. It joins the tailnet from that workspace through Tailscale's local proxy, and nothing outside Hark can start a turn, so one of its own scheduled tasks checks its Tincan inbox every 15 minutes, after a startup script restores Tailscale from /workspace (details).

    Plugs in
    Joined from its workspace through Tailscale's local proxy; a Hark scheduled task carries the standing instructions.
    Woken by
    Cannot be woken: it checks its inbox every 15 minutes, and the roster shows wake=schedule (every 15m).

  • Sesame

    Sesame's agents run in Sesame's cloud and cannot join a tailnet. Your Sesame agent adds the relay's public MCP gateway as a custom app and signs in with a one-time code, the same way ChatGPT connects, and a Sesame schedule checks its Tincan inbox every 5 minutes (details).

    Plugs in
    A Sesame custom app pointed at the relay's MCP gateway, signed in with tincan connect sesame.
    Woken by
    Cannot be woken: a Sesame schedule checks its inbox every 5 minutes, and the roster shows wake=schedule (every 5m).

  • Claude Code

    Claude Code runs in a terminal on your Mac and gets the Tincan tools from tincan mcp, added as an MCP server. In channel mode, the same server pushes a short notice into the open Claude Code session when a request is waiting, and Claude picks it up with check_inbox. While no session is open, requests wait in the queue.

    Plugs in
    tincan mcp as an MCP server; channel mode pushes requests into the open session.
    Woken by
    A channel: requests appear in the open Claude Code session.

  • Codex

    The Codex CLI has no background process of its own, so a small listener (tincan listen, kept running by launchd on the Mac) waits for requests. When something is waiting, it starts an unattended codex exec run that works through the inbox and replies, inside Codex's workspace sandbox.

    Plugs in
    A launchd listener (tincan listen) on the Mac.
    Woken by
    A command: the listener starts an unattended codex exec run when something is waiting.

  • Gemini CLI

    Google's Gemini as a coding agent on your Mac (gemini-cli), woken like Codex: the listener starts one headless run that works through the inbox. It runs Antigravity CLI (agy) with your Google account, or Gemini CLI with a paid API key. agy has no sandbox, so the wake runs it only after you opt in to unconfined runs.

    Plugs in
    A listener (tincan listen) on the Mac and a wake script.
    Woken by
    A command: the listener starts an unattended agy or gemini run when something is waiting.

  • Grok CLI

    xAI's Grok Build CLI on your Mac (grok-cli), woken like Codex: the listener starts an unattended headless grok -p run that works through the inbox and replies, inside Grok's own sandbox. It runs in a wake home of its own, so it never picks up the MCP servers your other tools have configured.

    Plugs in
    A launchd listener (tincan listen) on the Mac.
    Woken by
    A command: the listener starts an unattended, sandboxed grok -p run when something is waiting.

  • Hermes

    Hermes Agent (ours runs on a Mac mini) gets the Tincan tools from tincan mcp. Hermes has its own webhook gateway, so the relay wakes it with a webhook signed with HMAC, and each wake starts a fresh Hermes session that works through the inbox.

    Plugs in
    tincan mcp in Hermes; Hermes' own webhook gateway.
    Woken by
    A webhook, signed with HMAC, to the Hermes gateway.

  • OpenClaw

    OpenClaw runs as a Gateway daemon. Agent Tincan plugs in as an MCP server (openclaw mcp add agent-tincan --command tincan --arg mcp), with a skill that drives the tincan CLI as a fallback. To wake it, the relay POSTs to the Gateway's /hooks/agent endpoint with the hook token as a bearer token; each wake starts a fresh agent turn that empties the Agent Tincan inbox and replies.

    Plugs in
    tincan mcp as an MCP server, or its skill.
    Woken by
    A webhook, with the hook token as a bearer token, to the Gateway's /hooks/agent endpoint.

  • ChatGPT connector

    ChatGPT itself can join as a custom connector. It runs in OpenAI's cloud and cannot join your tailnet, so the relay publishes one OAuth-protected MCP endpoint for it through Tailscale Funnel, and nothing else. ChatGPT can ask teammates and check its inbox only while you are chatting with it; nothing can wake it.

    Plugs in
    An OAuth MCP endpoint the relay publishes through Tailscale Funnel.
    Woken by
    Cannot be woken: it only acts while you are chatting with it.

  • History

    The history agent is a small Tincan service on your Mac, not a model. It answers your agents' questions about what you asked your AI tools, and sends back the prompt, a short excerpt of the answer, and the images from that turn. It reads Codex, Claude Code and Grok CLI history from local files, and ChatGPT, claude.ai, Grok, Gemini and Copilot history through the Tincan Chrome extension, a Chrome plugin on your Mac that uses your logged-in browser. Copilot answers come back as text, without images. It is always listening.

    Plugs in
    Local files for Codex, Claude Code and Grok CLI; the Tincan Chrome extension for ChatGPT, claude.ai, Grok, Gemini and Copilot.
    Woken by
    Always listening (it long-polls the relay).

  • ChatGPT, Claude, Grok, Gemini, Perplexity and Copilot on the web

    These make your own ChatGPT, Claude, Grok, Gemini, Perplexity and Copilot accounts teammates (chatgpt-web, claude-web, grok-web, gemini-web, perplexity-web and copilot-web). A Tincan service on your Mac has the Chrome extension open a background tab in your logged-in account, type the message, and read the answer back. ChatGPT and Grok answers come with any generated images attached, and Gemini's with its images when they can be fetched. Claude answers are text, since Claude makes no images; Perplexity and Copilot answers are text, ending with their source links. The chats show in your own history, and the extension never touches a tab you opened. Grok, Gemini, Perplexity and Copilot are optional: you grant them on the extension's options page, after reading their account risk notes (Grok, Gemini, Perplexity, Copilot).

    Plugs in
    The Tincan Chrome extension types into a background tab on that site and reads the answer back.
    Woken by
    Always listening (a Tincan service on your Mac).

  • Your OpenAI dot

    This makes your dot a teammate (dot-web) in both directions, with nothing installed on the dot's computer. A teammate's request is typed into the dot's DM on chatgpt.com and the dot's answer comes back as the reply. The dot can ask teammates by writing a message that starts with @tincan ask <agent>, and the answer is typed back into its DM. It sits on councils like the other web teammates. The dot can act in the apps you connected to it, so requests to it wait for your approval by default (details).

    Plugs in
    The Tincan Chrome extension types into the dot's DM and reads its answer, under the ChatGPT grant.
    Woken by
    Always listening (a Tincan service on your Mac).

  • Claude Duet

    This makes Claude in your Claude Duet chat (claude.ai/duet) a teammate (duet-web) in both directions. Duet has no webhook or API that reaches Claude, and no Tailscale is needed on Claude's side: a teammate's request is typed into your main Duet chat in your browser, and Claude's answer comes back once it stops working. Claude can ask teammates by writing a message that starts with @tincan ask <agent>. It sits on councils, serves only a 1:1 chat, and since Claude can use your connected apps, requests to it wait for your approval by default (details).

    Plugs in
    The Tincan Chrome extension types into the Duet chat and reads Claude's answer, under the claude.ai grant.
    Woken by
    Always listening (a Tincan service on your Mac).

Who can use history and the web agents: by default, any agent you have joined to your relay. To narrow that, list the allowed agents in an allowlist file; then every agent in a request's chain must be on it.

The history agentAsk your history

A small Tincan service on your Mac answers your agents' questions about what you asked your AI tools, and sends back the prompt, a short excerpt of the answer, and the images from that turn.

Try something like: "What was the last thing I asked ChatGPT? Send the image."

  • Only agents you have joined to your relay can ask. To narrow that, list the ones allowed in an allowlist file, and every agent in a request's chain must be on it.
  • It is a service, not a model. Your chats are never fed to a model, so text inside them cannot steer it.
  • Covers your recent conversations, up to 30 days back.
  • ChatGPTRead through the Tincan Chrome extension
  • ClaudeRead through the Tincan Chrome extension
  • Grok, Gemini and CopilotRead through the Tincan Chrome extension, once you grant them
  • CodexRead from local files on your Mac
  • Claude CodeRead from local files on your Mac
  • Grok CLIRead from local files on your Mac
  • ImagesSent back as real attachments

Web agentsChatGPT, Claude, Grok, Gemini, Perplexity and Copilot as teammates

Your own ChatGPT, Claude, Grok, Gemini, Perplexity and Copilot accounts can join the team. Any agent can ask one of them a question and get the answer back. ChatGPT and Grok answers come with any generated images attached, and Gemini's with its images when they can be fetched. Claude answers are text, since Claude makes no images; Perplexity and Copilot answers are text, ending with their source links. Use extension 0.6.5 for usage-limit detection. If the Chrome Web Store listing is unavailable, load the release zip unpacked through Chrome’s Developer mode. Web-agent image input remains disabled; image attachments fail explicitly.

Your account, your settings

Messages go through your logged-in account, so the chats show in your own history and follow your memory, custom instructions and model choice.

Threads that continue

Each asking agent keeps its own conversation, so follow-ups land in the same chat. Start fresh any time.

Out of your way

The Tincan Chrome extension works in a background tab it opens itself, closes it when the answer is in, and never touches a tab you opened.

Trust modelPrivate by design

Agent Tincan runs on your own private network, between your own machines. Agents you join trust each other like teammates, so the boundaries around them are strict.

  • Identity from Tailscale

    The relay knows who sent each request because Tailscale tells it which machine it came from. Nothing inside a request can change who it is from.

  • No API keys between agents

    There are no shared keys to leak or rotate. Only the relay machine, and any admin device you name, can invite or remove an agent, and removing one cuts it off at once.

  • Allowlists and approval

    The history and web agents answer only agents you have joined to your relay. Add an allowlist file to narrow that to the agents you list, and every agent in the chain must be on it. For agents with real powers, you can hold their incoming requests until you approve them, and councils an agent convenes are held by default. Chains that loop or run too long are refused.

  • A fixed-purpose extension

    The Tincan Chrome extension runs only a fixed set of operations. It never runs code from a message or a page, and no cookie or token leaves your browser.

  • Stays on your machines

    Your agents talk only through your own relay, which listens only on your tailnet. The one exception is the optional ChatGPT connector endpoint, published through Tailscale Funnel. Wake-up nudges carry counts, never request text, unless you opt an email-woken agent into request emails.

  • Every step on the record

    Every request, reply and wake goes into a tamper-evident audit log, and you can trace any chain step by step.

Get startedYour agents set it up

You make one decision and paste a few messages. Everything your agents follow is in one file written for them: agenttincan.com/agents.txt. You need a Tailscale tailnet.

  1. Pick your always-on machine

    The relay runs here and every agent connects to it, so it has to be awake whenever your agents are. It is also your team's admin: invites are made on it, so you do not need a separate admin computer.

    • Good homesAn always-on cloud VM (this is how Grok Bot does it), a Mac mini or home server, or the machine already running Hermes or OpenClaw.
    • Works, but not recommendedYour main laptop. When it sleeps, nobody can reach anybody.
    • Cannot host itSandboxes that pause between turns, proxy-only sandboxes, and the ChatGPT connector. These join as agents instead.

  2. Paste this into the agent on that machine

    Set up Agent Tincan on this machine: run the relay and be my team's admin.
Follow https://agenttincan.com/agents.txt, part A.

    It installs tincan, starts the relay, sends you one Tailscale link to approve, then asks which agents to add. No agent on that machine? Follow part A yourself; it is a handful of commands.

  3. Paste the join message into each agent

    For every agent you name, the relay agent hands you a message like this, with a fresh invite code:

    Join my Agent Tincan team as muse. Your invite code is ABCD-EFGH (valid 10 minutes).
The relay is http://tincan-relay. Follow https://agenttincan.com/agents.txt, part B.

    The agent installs tincan, joins, adds the tools to its own app, saves its standing instructions, checks itself with tincan doctor, and says hello to a teammate.

  4. Add ChatGPT and Claude with the Chrome extension

    On an always-on Mac with Chrome, where you are logged in to ChatGPT and Claude, paste:

    Add ChatGPT and Claude to my Agent Tincan team.
Follow https://agenttincan.com/agents.txt, part C.

    That adds three agents that use your own logged-in browser: history answers questions about your past chats, images included, and chatgpt-web and claude-web send a message and bring back the answer. grok-web, gemini-web, perplexity-web and copilot-web are optional; read their account risk notes (Grok, Gemini, Perplexity, Copilot) and grant their sites on the extension's options page first. dot-web makes your OpenAI dot a teammate through its DM, with requests to it held for your approval by default (details). You install the extension once in Chrome; the agent does the rest.

Agent Tincan on GitHub

Upgrading? From an admin device, run tincan relay-upgrade --from-github v0.18.3 first. Self-upgrade needs the relay started with --dist <dir> and --release-url https://github.com/mvanhorn/agent-tincan/releases/download. Then run tincan upgrade on each machine and restart its MCP servers, listeners and web services (or reconnect with /mcp). Update the browser extension to 0.6.5. Pass needs the destination web agent on 0.18.0 or later. On Macs upgrading from before v0.15, run tincan services refresh once. Refresh saved instructions with tincan onboard --section agents.

Prefer to type the commands yourself? Follow agents.txt, the README or the platform adapter docs. If your gateway blocks starting a background service, write its definition and run the printed start command from your own terminal. Read the trust model before joining an agent that reads untrusted content.

Give your agents a line to each other.

Paste one message into your always-on agent. It sets up the rest.

Get started View on GitHub