AI agents (MCP)

Connect Claude — or any MCP client — to FounderDB so it can triage your inbox, reply and follow up from your mailboxes, set up your sequences, and keep your to-dos. Everything it does is recorded as the agent's, not yours.

Connect

claude.ai (and Claude Desktop / mobile)

1. In Claude, open Settings → Connectors → Add custom connector.
2. URL: https://founderdb.co/api/mcp. Leave the OAuth client on Use Claude's published identity (recommended) — nothing else to fill in.
3. Click Connect. You'll sign in to FounderDB, choose how the agent's actions are labelled ("Claude"), and allow access.

FounderDB supports OAuth 2.1 for MCP: Client ID Metadata Documents, dynamic client registration, PKCE (S256), and refresh tokens that rotate. Connections show in Outreach → Settings → AI agents, where you can disconnect them.

Claude Code

claude mcp add --transport http founderdb https://founderdb.co/api/mcp

Claude Code opens the FounderDB sign-in. Or skip OAuth with an access token from Outreach → Settings → AI agents (MCP) — it's shown once:

claude mcp add --transport http founderdb https://founderdb.co/api/mcp \
  --header "Authorization: Bearer fdb_…"

Other MCP clients

{
  "mcpServers": {
    "founderdb": {
      "type": "http",
      "url": "https://founderdb.co/api/mcp",
      "headers": { "Authorization": "Bearer fdb_…" }
    }
  }
}

The server speaks MCP over Streamable HTTP (stateless JSON responses) and supports protocol versions 2025-06-18, 2025-03-26 and 2024-11-05. A client that only supports stdio can use a bridge such as npx mcp-remote https://founderdb.co/api/mcp.

Getting new tools

New tools appear in your next new chat (claude.ai) or session (Claude Code, or /mcp → reconnect). There's no need to reconnect the connector.

What the agent can do

Try asking: "Go through my FounderDB to-do inbox, draft replies, and send the ones I approve." Or plan a sequence together — see the recommended sequence.

Editing sequences safely

• Every change is validated before it's saved; an invalid flow is never stored.
• A change that removes the step contacts are currently on is refused, with how many contacts it affects, unless the agent passes force.
• Removing or rewiring a step never silently deletes the steps after it — the agent has to ask for that with remove_cut_off.
• New sequences start disabled; deleting one needs its exact name. Steps are laid out automatically, so the builder shows the agent's changes neatly.

Sending email

• send_email really sends from your mailbox. If you've emailed the contact before, it continues that thread from the same mailbox; otherwise it needs a subject and starts a new one.
• Pass send_at (ISO 8601) to schedule a follow-up. Queued emails show in the conversation, marked as the agent's, until they go out, and can be cancelled.
• Contacts marked do not contact can't be emailed.

When something looks stuck

Emails, sequence steps and reply pickup are run by a background worker. Agents are told to call get_system_status when emails look late or a step didn't happen, instead of guessing. get_setup_overview and inspect_send_queue also include a short system status, and send_email, enroll_contact, resume_sequence_for_contact and run_quick_action add a note when the worker is down.

If the worker is offline, nothing is lost: what the agent sends or schedules is saved and goes out when it's back. The agent tells you that instead of retrying or re-sending.

Everything is on the record

• Every tool call — reads too — is logged in Outreach → Agent activity: which agent, what, when, which contact, what it sent, and whether it worked.
• Emails the agent sends are labelled "Claude · AI agent" in the conversation instead of "You". Its notes, stage moves, snoozes and variable changes are labelled too.
• To-dos it adds say "Added by Claude".