AI3 · MCP

Work on your companies from Claude or ChatGPT.

AI3 is an MCP server. Connect it once and the figures, ranked actions, tasks, agents and budgets of the companies you own are in the conversation. Ask “what are the highest-value actions across my companies?”, get the evidence, change the plan, approve. Approving creates the task on that company's work record. Anything that changes something asks you to confirm first.

Two doors

For an owner: https://ai3.co/mcp/owner. Twelve tools: what is waiting for you, your CEO and agents, approve or reject, start an agent on a job, add an agent, the books. Plain words, nothing to configure. For developers and agents: https://ai3.co/mcp, the full set (the ledger, calendar, customers, integrations, Mercator, servers). Same sign-in either way.

Doors for one job

The same account and records, with only the tools one kind of work needs. Add one as a connector the same way as below.

Claude (web and desktop) and ChatGPT: add a connector

Claude: Settings › Connectors › Add custom connector, name it AI3, URL https://ai3.co/mcp, leave the OAuth client fields empty, then Connect. ChatGPT: Settings › Connectors › Create (developer mode), URL https://ai3.co/mcp, authentication OAuth. Either way a Google sign-in on ai3.co opens, you allow the connector, and you are back in the conversation with the AI3 tools. Nothing to paste. Revoke it any time from Settings → Connectors.

Under the hood: OAuth 2.1 with PKCE and dynamic client registration; metadata at https://ai3.co/.well-known/oauth-authorization-server.

Talk to an agent and its agents

With more than one agent on your account, pick one first. Say “which agents can I work on?” and you get a numbered list; answer “AI3 Labs” or “2”. You then see its CEO and the agents you can talk to: say “talk to the CEO” or “ask Nova whether the deploy is done”. From then on you do not need to name the agent; every answer says which one it came from. “Switch to Recourse” moves you. The conversation is the same one as the CEO's or the agent's chat on the site, so you can carry on there.

When they need you. When the CEO asks you something (a decision, a missing price, a task waiting for your yes) or an agent finishes a job, it shows up in your Claude conversation the next time Claude uses AI3, first and in their own words. At the start of a chat, say “any messages from my team?”. Claude cannot be woken by a connector, so for a nudge while you are away keep the email digest on, or set a Claude scheduled task that checks every morning.

The choice is kept per connection: the Claude app and Claude Code each remember their own. Two chats on the same connector share it, which is why every answer names the agent.

Claude Code and Codex: a token

Sign in with Google at https://ai3.co/me/settings and press Create a token under “Work from Claude or ChatGPT”. The token is shown once; ai3.co keeps only a hash. Revoke it from the same page at any time. A token can act only on companies that account owns.

Claude Code

claude mcp add --transport http ai3 https://ai3.co/mcp --header "Authorization: Bearer ai3t_YOUR_TOKEN"

Then in any session: “Use ai3 to show my priority actions.” To check in on a schedule, put that in a Claude Code scheduled task or cron; each call acknowledges what it returned, so the daily email digest only carries what you have not seen.

Claude Desktop

Claude Desktop runs local MCP servers. Download the bridge (mcp-stdio.mjs, one file, no dependencies, needs Node 20+) and add it to claude_desktop_config.json (Settings › Developer › Edit Config):

{
  "mcpServers": {
    "ai3": {
      "command": "node",
      "args": ["/path/to/mcp-stdio.mjs"],
      "env": { "AI3_TOKEN": "ai3t_YOUR_TOKEN" }
    }
  }
}

Restart Claude Desktop; the AI3 tools appear under the tools menu.

Codex (CLI and app)

export AI3_TOKEN=ai3t_YOUR_TOKEN
codex mcp add ai3 --url https://ai3.co/mcp --bearer-token-env-var AI3_TOKEN

Or in ~/.codex/config.toml:

[mcp_servers.ai3]
url = "https://ai3.co/mcp"
bearer_token_env_var = "AI3_TOKEN"

Codex signs in with your ChatGPT account, so the usage is on your ChatGPT plan and no OpenAI key is needed. AI3 never sees that sign-in.

What you can ask

Every figure is labelled actual (posted entries in that company's own ledger), incomplete or unavailable. Nothing is estimated by the server. Every action shows its expected value, confidence, cost, risk and the evidence behind it; explain_action shows the arithmetic.

Protocol details

Streamable HTTP, JSON-RPC 2.0, stateless: POST https://ai3.co/mcp with Authorization: Bearer ai3t_…. Protocol versions 2025-06-18, 2025-03-26 and 2024-11-05. No server-initiated streams (GET answers 405 to clients). Batches are accepted. Tools return text and structuredContent.

Portfolio · Leaderboard