OpenAssistantAPI
→

Using OpenAssistant with agents

Connect Claude Code, Codex, OpenClaw, Meta Muse, Instinct, Grok Bot, or any agent to your assistant over MCP or the command line, so your agent can delegate real work to it.

Two ways to connect

Your assistant can take work from other agents. Pick the connection that fits the agent you use.

  • MCP server. Add one server URL to any MCP client. The client registers itself, sends you to a consent screen, and gets only the permissions you approve. See MCP server.
  • CLI and SDK. Install one npm package. Run it from a terminal, let any agent with a shell run it for you, or import the same client as an SDK in your own code. See CLI.

Set it up in your agent

Each section below ends with a block you can paste straight into that agent. It does the setup for you and tells you the one step it cannot do, which is approving access.

Claude Code

Add OpenAssistant as an MCP server, then approve access once.

  1. Add the server.
claude mcp add --transport http openassistant https://mcp.agnt.ai/agnt/mcp/openassistant
  1. Run /mcp inside Claude Code, pick openassistant, and sign in to approve access.
  2. Ask Claude Code to use it, for example: "Ask my assistant to find a time with Dana next week."

Paste this into Claude Code:

Connect OpenAssistant to this Claude Code setup. Run: claude mcp add --transport http openassistant https://mcp.agnt.ai/agnt/mcp/openassistant
Then tell me to run /mcp, choose openassistant, and approve access in the browser. Once I have, list the OpenAssistant tools you can see and give me one example of a task I could delegate.

Codex

Add the server to your Codex config, or let Codex run the CLI.

  1. Add this to ~/.codex/config.toml.
[mcp_servers.openassistant]
url = "https://mcp.agnt.ai/agnt/mcp/openassistant"
  1. Restart Codex and approve access when the sign-in screen opens.
  2. Prefer the command line? Install the CLI and tell Codex to run the openassistant command.
npm install -g @agnt-sdk/openassistant

Paste this into Codex:

Set up OpenAssistant for me. Add this to ~/.codex/config.toml, keeping everything already there:

[mcp_servers.openassistant]
url = "https://mcp.agnt.ai/agnt/mcp/openassistant"

Then tell me to restart Codex and approve access when the sign-in screen opens. If MCP is not available in this version, install the CLI instead with: npm install -g @agnt-sdk/openassistant

OpenClaw

Give OpenClaw the CLI. It can run it like any other command.

  1. Install the CLI on the machine OpenClaw runs on.
npm install -g @agnt-sdk/openassistant
  1. Sign in. You get a verification code by email, and this login gets its own revocable key.
openassistant login --email [email protected]
  1. Tell OpenClaw to use the openassistant command to delegate tasks, for example:
openassistant tasks create --assistant <assistantId> --message "Draft a follow-up to yesterday's call"

Paste this into OpenClaw:

Install the OpenAssistant CLI with: npm install -g @agnt-sdk/openassistant
Then run: openassistant login --email <my email>. I will give you the verification code from my inbox.
After that, use the openassistant command whenever I ask you to delegate work to my assistant, for example: openassistant tasks create --assistant <assistantId> --message "<what I asked>". Run openassistant --help to see what else it can do.

Meta Muse, Instinct, and Grok Bot

These personal assistants are built to take requests from you, not to add outside tools, so the dependable route is to let them reach your assistant the way a person would: by email or text.

  1. If your version can add MCP servers or run commands, use the paste block under "Other agents" instead.
  2. Otherwise, have the assistant write to your OpenAssistant assistant. Your assistant answers like a colleague and can do the work or ask you first.

Paste this into Meta Muse, Instinct, or Grok Bot:

I have an executive assistant on OpenAssistant. When I ask you to hand something off to it, email or text my OpenAssistant assistant at <my assistant's email or phone number> with the full request, any context it needs, and my name, then tell me what you sent.

Other agents

Anything that speaks MCP or can run a command works.

  1. Any MCP client: add this server URL in its settings, then approve access.
https://mcp.agnt.ai/agnt/mcp/openassistant
  1. Any agent with a shell: install the CLI and let it call openassistant.
npm install -g @agnt-sdk/openassistant
  1. Your own code: use the same package as a typed client.
import { PortalClient } from '@agnt-sdk/openassistant';
const client = new PortalClient({ apiKey: process.env.AGNT_API_KEY });
const { tasks } = await client.tasks.list();

Paste this into any agent:

I want you to be able to delegate work to my OpenAssistant executive assistant. If you support MCP, add this server URL and walk me through approving access: https://mcp.agnt.ai/agnt/mcp/openassistant
If you can run shell commands instead, install the CLI with npm install -g @agnt-sdk/openassistant, run openassistant login --email <my email>, and use openassistant tasks create to hand work over. Tell me which of the two you used.

Good to know

  • An admin turns MCP on. MCP is off by default. An account admin enables it once for the account.
  • You approve what it can touch. Each permission is a read or write pair for one area, like tasks, calendars, contacts, or drive. You can revoke any client at any time.
  • One key per login. Each CLI login mints its own labeled key, so use a fresh login for every machine and every agent.