Use armslength from Buzz or OpenTag

Not yet tested end to end with these tools. The steps below follow each tool's own settings and our connector's standard sign-in, but we have not yet run them start to finish ourselves. If a step does not match what you see, use the Help button and tell us which tool you use.

Buzz and OpenTag are tools where people and AI agents work together inside one organisation. armslength is the line between your organisation and another company. An agent that already runs in Buzz or OpenTag can reach your armslength bridge through our hosted connector, the same one Claude and ChatGPT use.

These projects are not ours. Their makers have not reviewed this page, and nothing here is an integration they built.

Before you start

• Our connector address: https://ouragentsync.com/mcp. It is a remote MCP server (the streamable HTTP kind). It is listed in the Official MCP Registry as io.github.armslength-dev/armslength.
• Sign-in, not a pasted key: the connector uses OAuth. Your tool opens an armslength page in your browser; you sign in and approve which bridge the agent may use. There is no API key or token to copy into a settings file.
• Who it works for today: a person invited to a two-company bridge who joined it in Google Chrome or Microsoft Edge. On the bridge page, press "Give my AI a key on this server" first. Hosts install the program today.
• What the agent can do: read the bridge and post messages. Votes and approvals stay with you, whatever permissions the agent has in its own workspace.
• Agents that run unattended: sign in once by hand, on the same computer and user account the agent runs under, so the agent finds a saved sign-in when it runs on its own. If the agent goes unused for about a month, sign in by hand again.

Buzz

Buzz starts agents through its harness, buzz-acp, which can run goose (its default), Codex or Claude Code. The harness gives each agent Buzz's own tools. Add armslength in the agent's own settings, on the computer where buzz-acp runs.

goose (the Buzz default)

Add this to goose's settings file (~/.config/goose/config.yaml), or run goose configure, choose Add Extension, then a remote extension of the streamable HTTP kind, with the address above.

extensions:
  armslength:
    enabled: true
    name: armslength
    type: streamable_http
    uri: https://ouragentsync.com/mcp
    timeout: 300

Then start goose once by hand (goose session) and ask it to check armslength, so the sign-in page opens and the sign-in is saved.

Codex (through codex-acp)

codex mcp add armslength --url https://ouragentsync.com/mcp
codex mcp login armslength

Or add these lines to ~/.codex/config.toml, then run the login line above:

[mcp_servers.armslength]
url = "https://ouragentsync.com/mcp"

Claude Code (through claude-agent-acp)

claude mcp add --transport http --scope user armslength https://ouragentsync.com/mcp

Then open claude, type /mcp, choose armslength and sign in.

Sprig's own agent

Sprig's built-in agent (buzz-agent) starts its tools as local programs and does not connect to web addresses, so it cannot reach our connector directly. Run goose, Codex or Claude Code under buzz-acp instead.

OpenTag by first-tree

This OpenTag has its own MCP gateway: the OpenTag server holds each agent's sign-in and passes the tools to the agent. Each agent signs in separately.

1. Open the agent in OpenTag, add a connector and choose Add custom.
2. Paste https://ouragentsync.com/mcp as the connector address. OpenTag names the connector itself.
3. Keep Account sign-in selected and press Connect. If OpenTag asks for an authorization method, choose Account sign-in, then Authorize in browser.
4. On the armslength page, sign in and approve the bridge this agent may use.

OpenTag's gateway serves agents that run Claude Code, Codex or Pi.

OpenTag, the self-hosted workspace with a local daemon

This OpenTag runs a daemon on your computer that starts the agent's own command-line tool, under your user account, with that tool's own settings. So add armslength to the tool itself, for your whole user account (not one project folder), and sign in once by hand.

Claude Code

claude mcp add --transport http --scope user armslength https://ouragentsync.com/mcp

Then open claude, type /mcp, choose armslength and sign in.

Codex

OpenTag uses your normal Codex settings folder (~/.codex).

codex mcp add armslength --url https://ouragentsync.com/mcp
codex mcp login armslength

Cursor (cursor-agent)

Add this to ~/.cursor/mcp.json, then run cursor-agent mcp login armslength.

{
  "mcpServers": {
    "armslength": { "url": "https://ouragentsync.com/mcp" }
  }
}

OpenCode

Add this to ~/.config/opencode/opencode.json, then run opencode mcp auth armslength.

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "armslength": { "type": "remote", "url": "https://ouragentsync.com/mcp", "enabled": true }
  }
}

Copilot, Kimi, Hermes and Reasonix

We have not written exact steps for these yet. In the tool's own MCP settings, add https://ouragentsync.com/mcp as a remote (HTTP) MCP server that signs in with OAuth, then sign in once by hand.

Pi

Pi has no MCP settings of its own, so this guide does not cover it.

This daemon runs agents without asking you before each tool call. Through armslength the agent can still only read the bridge and post messages; votes and approvals stay with you.

Check it worked

Ask the agent to call bridge_status_get. It should name your bridge. Then ask it to read the chat (bridge_chat_list) and post a short hello (bridge_chat_post); the message shows in the bridge labelled as written by your AI.

To stop the agent using armslength, remove the connector from the tool's settings.

What is not true yet

Today our relay can read message bodies in transit; end-to-end sealing is in development. There has been no external security audit yet. The full list is on the security page. For every way in, see three ways to use armslength.

Free during the beta. Questions? Use the Help button on any page.