Coming soon · preview

Bring your own agent.

Claude, ChatGPT, Claude Code, your own stack: point it at your Thread Otter project over MCP. Your agent gets the eyes and the hands. The judge, the pacing, and the ban guards stay ours, so nothing it proposes can burn your accounts.

Start free trial Request early access

the product works without any of this · the connector is for founders who already run their own agents

SOCKETone connector, any agent

One MCP server. Every agent speaks it.

A remote MCP endpoint with a project-scoped key. Add it to Claude or ChatGPT as a connector, to Claude Code as a server, or call it from any MCP client you run. No SDK, no webhook maze, no OAuth dance.

Claude + ChatGPT connect as a custom connector

Claude Code + agent frameworks connect as an MCP server

Keys are scoped to one project, revocable any time

Add to your agent

{
  "mcpServers": {
    "thread-otter": {
      "url": "https://api.threadotter.com/mcp",
      "headers": {
        "Authorization": "Bearer to_key_..."
      }
    }
  }
}

preview · final endpoint + key format may change before launch

HANDSwhat your agent can do

Read everything. Propose anything.

The same tools our own assistant works with, exposed to yours. Reads are instant. Writes are proposals: they land in the pipeline as drafts, never on a platform.

Reads

list_signals

scored buyer mentions across every watched source

read_thread

a full conversation with its draft + outcome history

list_leads

the person-centric directory: signals + cold imports

read_analytics

sends, replies, and outcomes per channel

read_context

the product brief + docs the drafts are grounded in

Proposals

propose_reply

draft a reply to a signal, in the founder's voice

propose_post

brief a post: the planner fans it out per channel

queue_draft

submit your agent's own text into the pipeline

add_context

feed a launch, doc, or changelog into the brain

illustrative names · the launched surface ships a stable, versioned tool set

RAILSwhy it can't hurt you

Your agent gets the hands, never the send button.

There is no raw send endpoint, by design. A badly prompted agent can waste its own rate limit; it cannot spam Reddit from your account. Every proposal goes through the exact pipeline autopilot uses.

That is the difference between an API that posts and an agent you would trust with your founder account.

Every write, same rails

Your agent proposesa reply, a post brief, or its own draft text

The judge reads itsame gate as autopilot: bot-tells and self-promo stripped, weak drafts rejected

Pacing schedules itper-account caps and human cadence, per channel

Then it sendsthrough your session or a managed account, with the receipt on the thread

judge · pacing · ban guards · receipts, on every proposal

DOORSways in, in order

MCP first. The rest follows.

Remote MCP server

the connector above: reads + proposals for any MCP client

ships first

Agent skill

install into Claude Code: our playbook baked in, so drafts follow strategy, not vibes

planned

CLI

scripts + cron: pull analytics, queue briefs from your pipeline

planned

REST API

the same surface as plain HTTPS for everything else

planned

included with the plan when it ships · no separate API tier

Your agent, our rails.

Want in early? Tell us what your agent should be able to do and we will build against real workflows, not guesses.

Request early access See what the teammate does

preview page · the connector is in design, the pipeline behind it already runs every day