← All features
MCP · Autopilot

Your AI agents can run the program.
Here's the working code.

Trcker exposes 12 MCP tools covering the full operator surface. A reference Claude Agent SDK implementation runs four loops against those tools — approve conversions, pause non-incremental partners, flag attribution mismatches, monitor caps. Fork it, deploy it, own it.

A first-class agent surface, not a bolt-on.

Trcker was built with agents in the architecture from day one. Every operator-facing action that exists as a dashboard click also exists as an MCP tool. No shimming, no screen-scraping, no "integration engineer" required.

R

7 read tools

get_partner_stats, get_offer_stats, get_conversions, get_fraud_summary, get_top_publishers, get_attribution_lift, get_incrementality_lift. Every operator question an agent could ask, answered in structured form.

W

5 write tools

create_offer, pause_partner, adjust_payout, approve_conversion, set_cap. Every write is audit-logged with a synthetic mcp:<brand> actor so your compliance trail cleanly separates agent-driven actions from dashboard actions.

A

Reference autopilot agent

A standalone TypeScript repo uses the Claude Agent SDK to run four loops against the tools: conversion triage, partner quality sweep, attribution rebalancing, and cap enforcement. Dry-run default, per-loop action caps, full JSONL audit log. Fork it, deploy it, own it.

P

Safe by design

Brand-scoped API key (agent can only touch one brand). Dry-run mode blocks writes until you opt in. Action caps prevent runaway loops. Protected-partner list for strategic relationships. 120 req/min rate limit on the MCP endpoint.

What the reference agent ships with

Four loops, each with its own system prompt, policy gate, and action cap. Exercising all 12 MCP tools in a single repo.

Conversion triage (every 15 min) — auto-approve low-fraud-risk pending conversions within policy
Partner quality sweep (nightly) — pause partners with statistically significant non-incremental lift
Attribution rebalance (weekly) — flag partners over-/under-credited by last-click vs multi-touch
Cap enforcement (every 5 min) — flag offers approaching their daily/weekly/monthly caps
Customizable — drop in your own loop in ~50 lines; the reference agent is a starting template, not a black box
Zero database, zero state — all state lives in Trcker via MCP. The only on-disk artifact is the audit log
Cost: ~$40-50/month of Anthropic API per brand at typical volume. Runs on your existing infrastructure.
Open source MIT — fork it, audit it, customize it. Your agent, your rules.

Trcker Autopilot vs. "we have an API"

Traditional platform
Trcker Autopilot
Agent surface
None — you integrate via webhooks and read the dashboard
12 MCP tools + a public reference Claude Agent SDK implementation
Transport
N/A
Remote Streamable HTTP at /api/mcp. Works with Claude Desktop, Claude Code, Cursor, any MCP client
Auth model
Session cookies or OAuth flow
Brand-scoped API key. One agent = one brand. No cross-brand access even if the key leaks
Audit trail
Dashboard action log (if any)
WorkOS Audit Log with synthetic mcp:<brand> actor. Clean separation of agent vs human writes
Writable
Varies — some platforms only expose reads to agents
5 write tools with MCP destructiveHint / idempotentHint annotations for client-side confirmation
Reference implementation
Build your own
Open-source autopilot agent with 4 loops, full audit, dry-run default

Frequently asked questions

What is MCP?

Model Context Protocol — an open standard from Anthropic for connecting AI agents to data sources and tools. Trcker's MCP server exposes 12 tools (7 read + 5 write) that any MCP-compatible agent can call. It's the agent equivalent of an API, but with standardized schemas, auth, and annotations.

Do I have to run the Autopilot agent, or can I use my own?

Your choice. The reference autopilot is one agent; you can point Claude Desktop, Claude Code, Cursor, or any MCP-compatible client at the same endpoint and get all 12 tools. The autopilot is there for people who want a turnkey running instance.

Will my agent do something destructive by accident?

The reference autopilot defaults to dry-run mode, has per-loop action caps (max 50 approvals per run, max 3 pauses per run), and supports a protected-partner list that the agent will never touch regardless of what the math says. You explicitly flip AUTOPILOT_MODE=execute after you've reviewed the dry-run logs. The MCP tools themselves carry destructiveHint annotations that well-behaved clients surface as confirmation dialogs.

What does it cost to run?

About $40-50/month of Anthropic API per brand at typical volume (around 20 partners, dozens of pending conversions/day). You pay Anthropic, not Trcker. Interval config tweaks can reduce this — e.g., triaging every 30 min instead of 15 min cuts that loop's cost by half.

Can I see the autopilot code?

Yes — it's open source (MIT license) once the reference agent goes public. Fork it, audit the hooks, customize the loops, extend it with your own policies. The repo is the canonical example of what Trcker's MCP surface is for.

What if I'm on Everflow / Impact / Partnerize?

They don't have an MCP surface at all (as of April 2026). Affise shipped a Desktop-only .dxt drop-in; it's read-focused. Trcker is the only affiliate platform where remote, writeable, agent-first is the default posture.

Related features

The first tracking platform where agents are the default.

MCP tools and the autopilot reference are included on every plan. Your brand API key ships agent-ready on day one.

Request Access