I built this on nights and weekends for myself. After it started working, I left the repo open.
Paste a GitHub URL. An agent clones the repo, opens files, follows imports and callers, and comes back with a wiki, an answer, a patch, or a PR review. It works from the tree in front of it — the same slow way you would if you had an afternoon free and enough patience to click around.
I started writing it after too many evenings of five tabs, a half-remembered grep, and a summary I still didn’t trust. The loop I wanted was smaller: read the tree, follow the callers, only write claims you can point at.
You can try it if that sounds useful.
A few shots from a live local run:
Wiki is the same family: paste a repo, get pages written from the real tree.
- Wiki — multi-page technical wiki from a repo URL
- Ask — questions with code-backed answers
- Code — give it a task, get a patch from a temporary worktree
- Review — pull request review / investigation
Bring your own model keys. They live in your browser storage and go out only with the run request, so the model can answer.
Live demo (when I’m keeping it up): rlmwiki.deepascii.com
You’ll need Bun and, for Agent mode, JCODE.
# optional agent runtime
curl -fsSL https://raw.githubusercontent.com/1jehuang/jcode/master/scripts/install.sh | bash
jcode --version
bun install
cp .env.example .env
# put a key in the UI under Keys, or export one for local CLI:
# export GEMINI_API_KEY=...
# export OPENAI_API_KEY=...
bun run server
# open http://127.0.0.1:3141CLI if you prefer a terminal:
bun ./bin/rlm-wiki.ts generate expressjs/express
bun ./bin/rlm-wiki.ts ask expressjs/express "How does routing work?"
bun ./bin/rlm-wiki.ts list
bun ./bin/rlm-wiki.ts serve --port 3141GitHub URL
→ clone / open workspace
→ agent reads real files
→ wiki / answer / patch / review
Under the hood there are two runtimes:
- Agent — JCODE (preferred for Code)
- RLM — vendored
rlm-bun(still used in some wiki / ask / review paths)
You don’t have to care about that on day one. Pick a model, paste a repo, press the button.
Open Keys in the top bar. Paste a provider key. Save.
Local browser storage only. Clear anytime. Env vars on the machine still help for CLI experiments; the web UI expects keys from the Keys panel.
rlm-wiki/
├── bin/rlm-wiki.ts CLI
├── src/ server + generators + runtimes
├── public/ web UI
├── vendor/rlm-bun/ legacy RLM runtime
├── docs/screenshots/ the pictures above
└── Dockerfile one-box deploy
Longer notes live in docs/. Changelog in CHANGELOG.md.
I run mine on Railway with the Dockerfile in this repo. Rough shape:
railway link
# BYOK in the browser for user runs
# optional: Cloudflare Access in front
railway upHealth check: GET /api/health.
Docker:
docker build -t rlm-wiki .
docker run --rm -p 3141:3141 rlm-wiki- GitHub is the main path.
- Agent mode needs JCODE on the machine that runs the server.
- A single wiki page can fail while the rest of the run keeps going.
- Hobby project. Sharp edges are normal. Issues and PRs help when something is broken for real.
I wanted a quieter coding loop — one that ends with a file path, not a vibe. Building that for myself was enough. Putting the code here is so you can fork it, host it, or steal one idea and put it somewhere better.
If you make something with it, tell me.
— MIT. Use it kindly.