Daedalus

The self-evolving, local-first coding assistant — powered by the Hermes Deep Mind engine.

---

Daedalus is a coding agent that runs on your own machine and your own models (Ollama local, or any of 24 providers). One engine, three surfaces — a rich terminal, a web IDE, and a VS Code extension — and a cognitive stack no other assistant ships: persistent memory, a failure "immune system", sleep-time compute, judge-verified goals, causal blast-radius prediction, MoE routing, and learned confidence calibration.

It doesn't just chat. It scaffolds runnable full-stack/mobile apps, verifies them, and prepares the deploy.

pip install daedalus-ai
daedalus            # rich terminal UI
daedalus web        # web IDE in your browser
daedalus app        # native desktop window

---

Why it's different

	Daedalus	Typical assistant
Runs on	your local models (Ollama) + 24 providers	1–3 cloud APIs
Memory	persistent across sessions; rebuilds context from checkpoints	forgets each session
Learns while idle	subconscious dreams/distills into memory & skills	only thinks when prompted
Stops when done	independent judge must confirm the goal	claims done optimistically
Before an edit	predicts blast radius from git co-change history	edits blind
Routing	validated-live MoE across providers, learned calibration	fixed model
Ships	scaffolds → verifies (eval gate) → deploys	code only

---

Install

pip install daedalus-ai                      # from PyPI
# optional extras:
pip install "daedalus-ai[app]"               # native desktop window (pywebview)
pip install "daedalus-ai[browser,desktop]"   # Playwright browser + PyAutoGUI desktop control

From source:

git clone https://github.com/KingLabsA/daedalus.git && cd daedalus
./install.sh

---

The three surfaces

1. Terminal — daedalus

A rich TUI: streamed tokens, markdown + syntax highlighting, inline diff-approve before writes, @file mentions, input history, Ctrl-C to cancel mid-run. Shows which provider each answer routed to.

Headless one-shot for CI / scripts / git hooks:

daedalus run "fix the failing test" --yes --json
# {"ok": true, "result": "...", "routed_to": "freellmapi", "files_changed": ["test_x.py"]}

2. Web IDE — daedalus web

One process serves the built UI + agent over a token-protected WebSocket. The Cockpit puts everything on one page:

• Chat with inline changeset review (accept/reject per file or per hunk)
• Monaco editor with live file tree
• Terminal (sandboxed command stream) + Preview (live app canvas)
• ShipBar — one-click Scaffold / Verify / Deploy

Plus a Mind dashboard: memory, subconscious activity, calibration curve, expert routing, blast-radius predictor, device doctor, model advisor.

3. Desktop — daedalus app

A real native window (pywebview → WebKit/WebView2/GTK), not a browser tab and not Tauri. ./build_app.sh produces a self-contained double-clickable Daedalus.app via PyInstaller (bundles Python + agent + UI).

4. VS Code — kinglabs.daedalus-vscode

Chat sidebar + native inline diffs driven by the changeset protocol.

---

Models — local-first, free-first

Daedalus auto-routes every request (easy → free local model, hard → strongest live provider) and only uses providers that actually answer — keys are validated with a live probe, not just checked for presence.

• Ollama (recommended): brew install ollama && ollama serve, then ollama pull qwen2.5-coder:7b. Local models are context-capped (HERMES_LOCAL_NUM_CTX, default 8192) and routed through Ollama's native API so num_ctx is honored (avoids an 8B model ballooning to 23 GB).
• FreeLLMAPI gateway: launch it, set FREELLMAPI_API_KEY.
• OpenCode Zen: OPENCODE_API_KEY (frontier models via one key).
• Free cloud tiers: Groq, Google AI Studio, Mistral, Cerebras, DeepSeek — put keys in .env.

daedalus doctor shows exactly which providers are live right now and what's missing. daedalus models shows which models your hardware can run.

Useful env: HERMES_AUTO_ROUTE=off, HERMES_LLM_TIMEOUT=120, OLLAMA_MODEL=<name>, HERMES_SUBCONSCIOUS=off.

---

The Deep Mind engine (Hermes)

• Context Engine — persistent memory (SQLite FTS), budgeted injection, checkpoint-based context reconstruction instead of truncation.
• Failure Immune System — every tool failure becomes a searchable "antibody"; before acting it checks whether it got burned this way before, in this repo.
• Subconscious — sleep-time compute: consolidates session experience into memory and distills repeated workflows into skills while idle.
• GoalJudge — an independent model verifies a /goal is truly complete before stopping (fail-open).
• Causal World Model — mines git co-change history + import graph to predict an edit's blast radius, and warns before high-risk writes.
• Model Orchestra (MoE) — classifies each task and routes to the best expert; committees + judged Max Mode best-of-N.
• Epistemic Engine — records predicted confidence vs actual outcomes; cost-aware routing driven by learned calibration.
• Senses — image analysis, video understanding (ffmpeg frame sampling), voice in/out.
• Native MCP client — connect any Model Context Protocol server via .hermes/mcp.json.
• Profile builder — first-launch interview pre-builds persona skill packs (developer, PM, doctor, engineer, …).

Self-* loop: self-learning (record → skill), self-verification (pytest gate), self-correction (errors feed back), self-implementation (writes & registers its own tools).

---

Text-to-app → verify → deploy

daedalus run "build a tailwind todo app called Tasks" --yes

Scaffold (17 kinds, runnable — not stubs): web · tailwind/shadcn · supabase · next/t3 · astro · svelte · api (FastAPI) · saas/fullstack · cli · mobile/ios/android (Expo) · mcp (a self-extending MCP tool server). Uses the canonical create-* generator when installed, else a built-in skeleton.

Verify (eval gate) — build must pass / code must compile / tests must be green / MCP must handshake. A broken app is blocked from deploying.

Deploy — detects the project, writes the provider config (vercel.json / netlify.toml / fly.toml / eas.json), and hands you the exact commands. Targets: Vercel, Netlify, Fly.io, Expo EAS. (It prepares everything; the authenticated login step is yours.)

---

Commands (CLI & web)

/goal · /multitask · /kanban · /memory · /remember · /dream · /distill · /subconscious · /blast <file> · /experts · /max · /route · /calibration · /see <img> · /say · /listen · /doctor · /models · /profile · /mcp · /ship <kind> <name> · /deploy <target> · /checkpoint · /safety · /provider · /reset — full list via /help.

---

Architecture

agent_ultimate.py        Think→Act→Observe loop, tool registry, safety
core/
  providers.py           24 providers, MoE-tier router, validated liveness, cost
  server.py              WebSocket protocol (chat, files, changesets, ship/deploy)
  context/               memory, immune system, checkpointer, budgeter
  cognition/             dream, distill, judge, subconscious
  intel/                 code intel, LSP client, embeddings, causal world model
  senses/                MoE orchestra, vision, voice
  epistemic/             calibration, cost router, Max Mode
  platform/              MCP client, doctor, profiler, model advisor
  changeset.py           per-file & per-hunk edit review
  scaffold.py            text-to-app (17 kinds)
  deploy.py / evalgate.py   deploy planning + pre-ship verification
desktop/                 React web UI (Cockpit, Mind, editor)
vscode-extension/        VS Code client
bench/swebench_runner.py SWE-bench harness

Standalone core/* packages import nothing from the monolith; the monolith re-exports for back-compat. 291 offline tests, CI on macOS + Linux × Python 3.11–3.13, ruff-clean.

---

Development

pip install -e ".[dev]"
python -m pytest tests/ -q          # offline suite (291)
python -m ruff check .              # lint gate
cd desktop && npm install && npm run build   # web UI

Contributions welcome — see CONTRIBUTING.md. Security policy in SECURITY.md.

---

Status & honest notes

• Positioning: single-user, local-first by design (not multi-tenant/SaaS).
• Desktop: the pywebview daedalus app is the recommended standalone; the legacy Tauri path is parked on an upstream macOS bug.
• Benchmarks: a SWE-bench harness ships in bench/, but no score is published yet — run it locally with a strong provider before quoting one.

License

MIT © KingLabs. Product: Daedalus. Engine: Hermes.