The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Trail listing page.
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.
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):
On Windows, write the real path instead of ~: the registration stores the argument verbatim
and Node does not expand it.
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.
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.
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/settings.json.~/.gemini/settings.json.~/.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:
Clients without a hook mechanism are covered too: the server's own instructions tell the agent to fetch the summary itself before acting.
Point it at your server once and it stays pointed:
Without that variable, login goes to the hosted service.
The token is stored at ~/.config/trail/token.json (mode 600). Revoke it whenever you want from
the "AI token" panel in the app.
| Variable | What it does |
|---|---|
TRAIL_API_URL | Your own Trail's address. Saved after the first login. Defaults to the hosted service. |
TRAIL_PROJECT | Forces the project name (default: the open folder). |
TRAIL_API_TOKEN | A token directly, skipping the browser login. Useful in CI. |
TRAIL_LANG | Language of what the connector prints on screen: pt or en. Without it, your system's language, and English when your system states none. Never changes what the agent reads. |
The former names (TETHER_*) keep working, with no deadline - the product used to be called
Tether, and nothing anyone already configured has to change.
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.
Node 18 or newer.
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.