本論文の製品化版 Scalable Behaviour Cloning Using via Skill Distillation(arXiv:2606.32014)を使ったブラウザ上のスケーラブル・ビヘイビア・クローニング

各ウェブサイトに1つのバケツが蓄積される→、自分のブラウザのタスクを記録してください 能力**、そしてすべての能力は一つの再利用可能なスキルに凝縮→使えます Claude Desktop(およびClaude Code)で。シングルユーザーのローカル製品。あなたは 自分専用のLLM APIキー。ランタイムでは純粋なPython(Nodeは不要)。ベンチマークではありません。

各ウェブサイトは、例えば github.com/{login-with-credentials, signup-with-email, …} — 新しい録音は セグメントごとに原子化され、各セグメントは能力とセグメントに分類されます。 同じ能力が一つのバケツに集まり、それが一つのスキルに凝縮されるのです。

レイアウト

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//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/-/SKILL.md (auto)
                 └─ Claude Desktop: .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

State lives under data/harness/: buckets.json, registry.json, and skills///{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 . 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.