The agent is the AISWEC GitHub App and three workflows that plan, build and answer review inside your repository. They run on your own GitHub Actions, with your own model key, and every line of them is public.
FSL-1.1-ALv2 · Apache 2.0 two years after each release
.github/workflows/aiswec-refine.yml
calls refine
.github/workflows/aiswec-build.yml
calls build
.github/workflows/aiswec-revise.yml
calls revise
.github/aiswec.yml
maintainers, checks, model
CLAUDE_CODE_OAUTH_TOKEN
your model key, as a secret
aiswec
the label that assigns a task, and five more
Each stage is a reusable GitHub Actions workflow in aiswec/agent, started by something you do on GitHub. Unless you say otherwise, nothing is built before you approve a plan, and nothing ever merges itself.
Write the task as you would for a colleague, then add the aiswec label. Opening an issue alone starts nothing.
Reads the issue, the repository and the vendor's docs at the version you use, then posts a plan. If your answer would change the result, it asks first, with options.
Reply approve, lgtm or ship it. The plan is built as you saw it, and one edited after your approval needs a new one.
Writes the change on a branch of its own, runs your checks and opens a pull request. Still failing after three tries, it opens a draft that says why.
Comments, inline notes, whole reviews: review it as you would a colleague's, for as many rounds as you like.
One reply per round, with commits on the same branch when you asked for a change. Never a second pull request.
Always yours. No stage has a step that merges, so the change lands when you say it does.
Pause a task with the aiswec:hold label, or take the aiswec label off and nothing new starts. Any stage can be run again from your repository's Actions tab.
How it works, step by stepThe same stages either way, on GitHub's runners in your repository, where your code is checked out and nowhere else. What differs is who sets the repository up, and whose name the work carries.
@v2, and a settings file that names your
maintainers and your checks.
@v2 to read every change
before it runs in your repository.
Either way the license is FSL-1.1-ALv2: read, run, change and self-host the agent for your own work and for your clients, but do not offer it as a competing commercial service. Each release becomes Apache 2.0 two years after it ships, and the license itself is what counts, not this summary.
Any AI agent that reads text and holds a key can be targeted by a prompt injection. The security manual says plainly which protections the workflow enforces, which the model is only told to keep, and what to turn on in GitHub to close the rest.
Short, plain and the same as the docs folder of aiswec/agent, so what you read here is what ships with the code.
.github/aiswec.yml.
agent/docs/settings
All of them start at aiswec.com/agent/docs. Suite, the hosted app, has a manual of its own.
Install the App and Suite sets the repository up for you, or copy three files and run the agent on your own. Your first project in Suite is free, and the code is yours to read, run and self-host.