Teamtrail
You pick the destination. Your agents walk the trail.
teamtrail.app · Docs · Try it: npx teamtrail@next
Teamtrail is a local trail map for planning and executing big work with your AI coding agents. You describe a trail; your agents chart it as decision tickets, work through them, and bring the calls that need you to a single inbox. Scout, the Lookout's trail lead, briefs you on what changed while you were away and keeps spend and token usage in view.
- Trails and decision tickets: a live map of what's decided, in progress, blocked and out of scope.
- Waiting on you: every question from every agent, in one inbox, answered in-window.
- Decision records: each call is written down with its alternatives, rationale and transcript.
- Live spend: per-session and per-trail tokens and cost.
- Your data stays yours: trails are plain markdown in
./.teamtrail/ in your repo. Nothing leaves your machine unless you connect Slack.
Works with any coding agent that connects over MCP. Claude Code and Codex are the primary supported agents.
Why Teamtrail
Your agents can already write the code. What's missing is a place to run them like a team: one where you set the direction, they do the work, and you make the calls that matter.
- Bring any agent. Everything runs through MCP tools. Claude Code today, Codex tomorrow, both at once. Teamtrail never calls a model itself and never locks you into a vendor.
- Local-first, for real. Trails live as plain files in your repo's
.teamtrail/ folder, or in the GitHub Issues or Linear tracker you already use. Nothing needs a Teamtrail server. Slack is optional: its questions, answers and digests pass through Teamtrail's Scout service, but your trails, decisions and approvals never leave your machine.
- A deterministic core. Lifecycle, state and ranking are code, not prompts. The same inputs always give the same result, so you can trust what the map tells you.
- Evidence over claims. Agents bring the decisions that matter to you, with the alternatives and the reasoning, and you decide. Less babysitting, not less judgment.
- Not another activity feed. Teamtrail isn't a prettier log of what agents did. It's where work gets steered.
Features
Plan and steer
- Trails and decision tickets: research, prototype, junction, task, execute and review tickets with blocking edges, on a live map.
- Discovery first: by default, agents write a brief you accept before implementation starts, and each ticket's acceptance criteria need evidence when the work is done.
- Plans that adapt: a trail lists the assumptions it relies on. When an agent finds one is wrong, it shows the evidence and proposes a new route; nothing changes until you approve.
- Project knowledge: what agents learn about your product is kept as entries you confirm, committed so your team shares them.
Only what needs you
- Waiting on you: one ranked list of questions, approvals, blocked work, briefs and new plans, each with the reason it's there. Focus for an hour, set quiet hours, and the rest waits in a digest.
- Risk policy: you decide what always needs you (database changes, auth, billing, deleting data, new dependencies). Agents can't make those calls on their own.
- Away briefing: Scout tells you what changed while you were away.
- Scout in Slack: answer agents' questions and approve work from Slack, and ask Scout things like “what's blocking #12?”, “what did we decide about retention?” or “who's working?”. High-risk calls are answered in the Lookout only.
Many agents at once
- Separate workspaces: each claimed ticket can get its own git worktree and branch. Agents say which files they'll change, you're warned before two collide, and a ticket to bring the work together is one click away. Teamtrail never merges or pushes for you.
- Model matching: per trail, each agent gets the tickets that fit the model it runs on, from research on fast models to junctions on the strongest. Teamtrail never picks or switches a model.
- Guides: a ticket can ask for a reviewer, QA, security, docs or architect guide, with a checklist the agent must answer before the work counts. Add more from a moderated catalog, or write your own with
teamtrail guides new.
- Remote agents: agents on other machines or in cloud sandboxes join over HTTP (
teamtrail mcp --http), each with a token you can revoke.
Evidence you can trust
以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗
