Introduction

Teamtrail is a local trail map for big work done by AI coding agents: you describe a trail, your agents map and do the work, and you make the calls that matter.

#How it works

1. You describe a trail, usually from your agent's chat with /scout new "<idea>".
2. The agent frames it with you, studies the repo, and charts a trail map: a set of decision tickets wired with blocking edges.
3. Agents claim tickets over MCP, ask you questions, record decisions with evidence, and prove their work against the checks and criteria you agreed on.
4. You steer in the Lookout, a web app that teamtrail serves on your machine, or straight from the chat with /scout.

#What matters for developers

• Local-first. Trails are plain files in your repo's .teamtrail/ folder, or issues in your GitHub or Linear tracker. There is no Teamtrail server.
• Agent-agnostic. Everything goes through MCP tools. Claude Code and Codex are the primary agents, and any MCP client works.
• No model calls. Teamtrail never calls a model. Agents write briefs, knowledge and criteria; Teamtrail gives them structure and enforces the gates in code.
• Deterministic. Lifecycle, gates, ranking and selection are code, not prompts. The same inputs give the same result, and every automated choice comes with a reason.
• You decide. Teamtrail never answers a question or approves anything on your behalf. Agents may make small calls themselves when your risk policy allows it, and every one is recorded for you to review.
• Not an executor. The local product doesn't start or schedule agents. It can prepare git worktrees for agents that ask.

#Where to go next