ForgeTrail gives your next coding session a place to start. It keeps the phase, decisions, and handoff in your repository so an agent can read them when you resume.

Start with Lite for a small project. Add the full templates or MCP when you need them. You do not have to run all seven phases to try the method. ForgeTrail ships as files your agent reads and follows, plus an optional installer, MCP server, and hooks. It is not a library or a SaaS.

Two sessions

A fictional example: desk-stamp, a weekend CLI that stamps today's date on markdown notes.

Session one. You write a short spec and paste the kickoff line. The agent drafts the Phase 1 brief, asks about the gaps, and waits. You approve it. The agent records the phase, the decision to keep everything on disk with no database, a Windows install gotcha, and a note for next time: agree the first stamp command, then scaffold.

Session two, days later. You open a fresh chat in the same folder and say "continue." The agent reads the ledger first. It does not reopen the database question. It asks only about the first stamp command, the last open Phase 1 item. Once you agree, it asks to move to Phase 2 and follows that phase's guidance: scaffold the app, wire the main flow end to end, and add a passing verify script.

What the agent writes

The record is appledger/ in your project: profiles/forgetrail.yaml for the phase, and records for decisions, lessons, and the session handoff. The ForgeTrail repository is the template source, and its backlog stays in RESUME.md and TODO.md. Application repositories carry the ledger. A shortened excerpt from the end of session one:

phase: plan
status: in_progress
choice: TypeScript CLI, files on disk only, no hosted store
left_off: Phase 1 brief approved. The first stamp command is not agreed.
next_steps:
  - Agree the first stamp command, then scaffold.

ForgeTrail instructs the agent to read that ledger, preserve approved decisions, and pause at approval gates. The agent must write the handoff. Those updates are not automatic. Optional hooks point a new session at appledger/ and remind it to update the session record. The full example is on Continuity.

One system, three parts

ForgeTrail guides the work through phases with approval gates. AppLedger keeps the record of that work in appledger/. xFacts labels describe what was built. Each works on its own.

Start with Lite

The recommended first path is Try: write a what-not-how spec, drop in ForgeTrail Lite, and paste one kickoff line. No Node and no MCP required if you copy the Lite file by hand.

You do not have to run all seven phases. Lite is enough for a small tool. The shortest supported first task is: create appledger/, draft docs/PHASE_1_BRIEF.md, and wait for you to approve that brief before any application scaffold.

Then the installer and MCP

PathWho uses itWhat it is
LiteAnyone with a coding agent that reads filesOne protocol file. The agent writes appledger/.
CLI (forgetrail)People with Node.js 20+ who want files placedInstaller. Writes Lite and Cursor hooks, or the full template tree. Skips files that already exist. Does not run the agent.
MCP (forgetrail-mcp)Cursor or Claude users who want tools in the IDEPhase guidance, templates, and lessons search. The ledger still lives in the app repo.

Do not add forgetrail to an app's dependencies. The trail is the files in your project.

Lessons shipped in Lite and in MCP search are first-party notes from Catalyst Forge production work. They are not independent adoption evidence.

Flags, the phase table, and install details live in the docs.

Open source under Apache License 2.0. Built by Catalyst Forge.