Trail

usetrail

The MCP server for Trail - it gives your coding agent your team's tracker and the project's durable memory, scoped to the folder you have open.

Your agent stops working blind. It reads what is open, writes what it did, and records what it learned the hard way - in the same board your team looks at, not in a scratch file that only that session can see.

This package is the client. It talks to your Trail account over the REST API; there is no database here. Dependencies are @modelcontextprotocol/sdk, zod and the built-in fetch.

Quick start

# 1. connect this machine to your Trail account (opens your browser; no token to copy)npx -y usetrail@latest login
# 2. register it with your agent - this example is Claude Codeclaude mcp add trail -s user -- npx -y usetrail@latest

No account yet? Create one at usetrail.dev - the free plan covers three projects and three people.

Not tied to one agent. This is a plain MCP server over stdio, speaking the three protocol versions in use today. Any MCP-compatible client can run it; only the registration command differs, and each client documents its own.

Prefer a local checkout? Clone it instead, and the connector keeps itself up to date on its own (see Updating):

git clone https://github.com/arsprengel/trail-mcp.git ~/.trail-mcpnpm --prefix ~/.trail-mcp install --omit=devnode ~/.trail-mcp/bin.js loginclaude mcp add trail -s user -- node ~/.trail-mcp/bin.js

On Windows, write the real path instead of ~: the registration stores the argument verbatim and Node does not expand it.

What your agent gets

The tracker - list_items, get_item, add_item, update_item, move_item, get_next, delete_item. Real items on your team's board, with type, status, priority, assignees and history. Work the agent picks up shows as in progress and closes when it is done, so the board is not a story someone has to retell.

Project memory - list_memory, get_memory, add_memory, update_memory, review_memory. The durable knowledge of a project: the deploy that has a trap in it, the decision nobody should relitigate, the command that is not in any README. Shared by everyone on the project, human or agent. A new session reads it before acting instead of rediscovering it.

Reminders - add_reminder, list_reminders, update_reminder, delete_reminder. Things with a date that must not be silently missed - and the agent closes or reschedules them itself once the date arrives.

Attachments - add_attachment, get_attachment. Files that belong to an item.

Which project it writes to

Trail holds many projects. The connector uses the folder you have open as the project - the folder name, or whatever a .trail file next to it says, or the TRAIL_PROJECT variable. So the agent writes to the right project without being told, and two repositories never bleed into each other. To reach across, the agent passes project explicitly on the tool.

Session hook (Claude Code, Gemini CLI, Antigravity)

Installed for you, once, when you log in - and on the connector's first run, which covers people who paste a token into the registration instead of logging in through the browser. Every session that starts in a project with a tracker gets the open items and the project memory in context from the first message - the agent starts knowing, instead of depending on someone remembering to ask. Fail-silent by design: no login or no network and the hook stays quiet.

Each tool keeps this in its own place, so the connector writes to the one that tool actually reads:

• Claude Code - a session-start hook in ~/.claude/settings.json.
• Gemini CLI - a session-start hook in ~/.gemini/settings.json.
• Antigravity - the tools are registered in ~/.gemini/config/mcp_config.json, which is the file it actually reads: gemini mcp add does not connect Trail in Antigravity, because that command writes to ~/.gemini/settings.json instead. Field-tested on Antigravity 2.11 (Windows), of the three things that live in that config folder only the tool registration is honored - the opening-summary hook is not run, and a plugin folder placed exactly where its own documentation says is not read. So on Antigravity the agent gets the tools and the server's working rules, and fetches the project summary itself on the first turn instead of receiving it. The plugin path is built and tested behind a switch, waiting for that to change.

usetrail status tells you, per tool found on this machine, whether it is on - so you never have to go digging through settings files to find out.

It never fights you. A tool that isn't on the machine gets nothing created for it, and hooks install will not fabricate a config for a program you don't have. If you remove it with hooks uninstall, it does not come back on its own - put it back with:

npx -y usetrail@latest hooks install

Clients without a hook mechanism are covered too: the server's own instructions tell the agent to fetch the summary itself before acting.

Your own Trail

Point it at your server once and it stays pointed:

TRAIL_API_URL=https://your-trail npx -y usetrail@latest login

Without that variable, login goes to the hosted service.

Commands

npx -y usetrail@latest            # start the MCP server over stdio - what your agent runsnpx -y usetrail@latest login      # connect this machine (browser confirmation, no token copying)npx -y usetrail@latest status     # server address, current project, whether a token is presentnpx -y usetrail@latest logout     # delete the saved tokennpx -y usetrail@latest doctor     # find the install on this machine and unstick auto-updates

The token is stored at ~/.config/trail/token.json (mode 600). Revoke it whenever you want from the "AI token" panel in the app.

Configuration

The former names (TETHER_*) keep working, with no deadline - the product used to be called Tether, and nothing anyone already configured has to change.

Updating

The @latest is not decoration. Registered as npx -y usetrail@latest, the connector resolves the published version every time your agent starts a session, so a new release reaches you without anyone doing anything. Registered as bare npx usetrail, npm reuses whatever copy it cached the first time - and that machine stays on that version forever. Use @latest.

The cost of @latest is about a second at session start, and a network call. Neither matters in practice: the connector talks to Trail over the network anyway.

A cloned install updates itself differently: on startup it fires a quiet fast-forward pull in the background, at most every six hours. The running session keeps the version it loaded; the next one starts updated. No network or a local conflict and nothing happens, silently.

Stuck on an old version? npx -y usetrail@latest doctor finds the install, reports the version and unsticks it.

Requirements

Node 18 or newer.

License

MIT. This package is the client - the connector you run on your own machine. Use it, fork it, adapt it. The Trail service it talks to is a separate, hosted product.