Journey Forge Local

Productized companion to our paper Scalable Behaviour Cloning on Browser Using via Skill Distillation (arXiv:2606.32014). See Paper & citation.

Record your own browser tasks → each website accumulates a bucket of capabilities, and every capability distills into one reusable skill → use them in Claude Desktop (and Claude Code). A single-user, local product. You bring your own LLM API key. Pure Python at runtime (no Node needed). Not a benchmark.

extension (record free-form task)
   └─► localhost server :8099   /v1/traces/init → chunks → finalize
          └─ assemble → data/traces/<id>/trace.json  (intent + events)
          └─ background harness pipeline:
               atomize → classify (capability) → bucket (per domain+capability)
               → distill bucket → SKILL.md + TRACE_GUIDE.md  → install
                 ├─ Claude Code:    ~/.claude/skills/<domain>-<capability>/SKILL.md (auto)
                 └─ Claude Desktop: <name>.zip → upload via Settings → Skills
   /api/buckets  exposes per-site capability buckets + skills

Each website is a folder of capability buckets, e.g. github.com/{login-with-credentials, signup-with-email, …} — a new recording is atomized into segments, each segment classified into a capability, and segments of the same capability pooled into one bucket that distills into one skill.

Layout

Path	What
entry/main.py	Desktop entry — starts the server, opens the panel in your browser
app/dist/index.html	Control panel (zero-build, served by the server)
extension/	The recorder (product fork: points at localhost, free-form only)
server/server.py	Local ingestion + control API
harness/	Distillation pipeline: atomize / classify / bucket / distill / install (pure Python)
harness/main.py	CLI: ingest / distill / status / consolidate / query
scripts/start.sh	Headless dev launcher
docs/	Trace schema + Claude Desktop setup

State lives under data/harness/: buckets.json, registry.json, and skills/<domain>/<capability>/{SKILL.md,TRACE_GUIDE.md,meta.json,evidence.jsonl}.

Quick start

1. Configure & run the server

   cp config.example.env .env.local      # then set SF_LLM_KEY=sk-ant-...
   ./scripts/start.sh                     # or: python entry/main.py  (native window)

   The panel is at http://127.0.0.1:8099/. The server seeds a default key (jfl-local-dev-key) the extension ships with, so it connects automatically.

2. Build & load the extension

   cd extension && pnpm install && pnpm build

   Chrome → chrome://extensions → enable Developer mode → Load unpacked → select extension/dist/chrome-mv3. It's pre-pointed at the local server.

3. Record → finalize. Record a short task with the extension, stop, label, upload. With auto-distill on, a skill appears within ~1–3 min:

   • Claude Code: already installed under your skills root.
   • Claude Desktop: download the .zip from the panel (Trajectories) and upload it in Settings → Skills. See docs/claude-desktop-setup.md.

4. (Optional) real browser execution. Panel → Browser execution → Configure Playwright MCP so Claude Desktop can actually click/type/navigate. Restart Claude Desktop after.

Notes

• Skills don't grant tools. A SKILL.md is injected instructions. To execute its steps in a browser you must configure a browser MCP (Playwright) separately — the panel automates this for Claude Desktop.
• LLM: the distiller speaks the Anthropic Messages API natively (default claude-opus-4-8). Point SF_LLM_BASE at an OpenAI-compatible gateway to use that path instead.
• Data lives under data/ (git-ignored). Nothing leaves your machine except the distillation calls to your configured LLM.

Paper & citation

This tool is the productized companion to our research on distilling reusable browser skills from human interaction traces — turning scalable human browsing into a library of per-site, per-capability skills that browser agents can reuse:

Scalable Behaviour Cloning on Browser Using via Skill Distillation Kaisen Yang, Zheng Jiang, Yuzhao Peng, Houde Qian, Boshi Zhang, Youjie Zheng, Shijin Hong, Qingle Liu, Ruoyu Han, Bohan Lyu, Bingxiang He, Eren Cai, Calvin Xiao, Qinhuai Na. arXiv:2606.32014, 2026. https://arxiv.org/abs/2606.32014

If you use this project or the ideas behind it, please cite:

@article{yang2026scalable,
  title   = {Scalable Behaviour Cloning on Browser Using via Skill Distillation},
  author  = {Yang, Kaisen and Jiang, Zheng and Peng, Yuzhao and Qian, Houde and
             Zhang, Boshi and Zheng, Youjie and Hong, Shijin and Liu, Qingle and
             Han, Ruoyu and Lyu, Bohan and He, Bingxiang and Cai, Eren and
             Xiao, Calvin and Na, Qinhuai},
  journal = {arXiv preprint arXiv:2606.32014},
  year    = {2026}
}