Getting started

One binary, one config directory, your git repos. Forge is currently a personal research system — the flow below is exactly how it runs here, and what a release would look like.

Build and start

$ git clone --recurse-submodules git@github.com:Ronin11/forge.git
$ cd forge && go build -o forge ./cmd/forge
$ ./forge daemon start

First boot bootstraps everything under ~/.forge: the SQLite database, a worker config, an auth token, and — the interesting part — your fork of the base directives library, seeded with the standard roles, the flow pipeline workflows, and starter routines. The daemon spawns its own local worker. The web UI is at http://127.0.0.1:7340.

Point it at a repository

$ curl -X POST http://127.0.0.1:7340/api/v1/repositories \
    -d '{"path": "~/Projects/myapp"}'

Any local git checkout with an origin works; the Repos page does the same with a "+". Drop a forge.toml in the repo declaring its checks (Forge will propose a skeleton if you don't) — those are what L1 verification re-runs on every attempt.

Give it work

# a small, direct task — one agent, verified, done
$ forge task add "tighten the flaky retry test in pkg/client" --repo myapp

# a real ask — planned, fanned out, supervised, merged
$ forge task add "add CSV export to every report page" \
    --repo myapp --mode plan --size M --integrate --autonomy auto

The second form is the whole machine: a planner decomposes the ask into a task DAG, builders run in parallel under write-set leases, every attempt is verified (L0–L3), green work goes through the merge queue, and a supervisor judges the settled batch — repairing, or accepting with scores.

Watch it run

The Forge dashboard during a benchmark run — **The dashboard mid-benchmark:** budget windows against their targets, the attempt timeline, every repository's state, and the attention queue — including a failed run, because the UI never flatters.

Useful commands while it works:

$ forge task show <id>        # state, attempts, verification trail
$ forge queue                # what's admitted, what's blocked and why
$ forge directives search verify   # the library, from the CLI
$ forge bench list rebuild-equitizr  # the benchmark's scored history

Make it yours

Edit the library — ~/.forge/directives is a git repo; the daemon hot-reloads within 30s and keeps the last good tree if you break it. forge directives update merges base-library improvements.
Schedule routines — a routine is a trigger (cron + envelope) pointing at a directive, script, or workflow. The seeds ship disabled; enable what you want.
Set the budget — admission targets, hard stops, and quiet hours in ~/.forge/config.toml; the dashboard shows the windows live.

Honest footnote: Forge is built and operated by one person, largely by the agents it orchestrates. If you're reading this as an engineer evaluating the ideas rather than a user installing a product — that's the intended audience, and the architecture docs are for you.