Local-model mode for the OpenClaw/Codex setup you already use.
Clawdeck is not another agent framework and it is not a separate worse Codex clone. It is a thin operating layer that adopts your existing OpenClaw workspace, makes the active model path local-first through Ollama, and gives both OpenClaw and Codex a shared CLAWDECK.md local-mode contract.
Once OpenClaw, Ollama, and the model weights are installed, local file/code work can keep running without wifi. Internet search, remote APIs, account auth, cloud models, and some tool-heavy workflows still need network access.
Adopt the OpenClaw workspace you already have:
npx @nicdunz/clawdeck setup --yesThat one command adopts the workspace, writes the backed-up local OpenClaw profile, runs the readiness drill, and prints the next command.
Manual path:
npx @nicdunz/clawdeck adopt
npx @nicdunz/clawdeck apply --yes
npx @nicdunz/clawdeck drill
npx @nicdunz/clawdeck smoke
npx @nicdunz/clawdeck handoffOr point it at a specific project/workspace:
npx @nicdunz/clawdeck setup . --yes
npx @nicdunz/clawdeck adopt .
npx @nicdunz/clawdeck apply --workspace . --yes
npx @nicdunz/clawdeck auditIf the scoped npm package is not published in your registry yet, use the GitHub installer:
npx github:dicnunz/clawdeck adoptYou get:
- a one-command first run through
clawdeck setup --yes CLAWDECK.md: the local-mode switchboard for OpenClaw and Codex- a non-destructive
AGENTS.mdpointer so existing agents see the local contract - a backed-up OpenClaw config apply
- active Ollama defaults without deleting existing providers/plugins/auth
- no hosted model fallback in the active model path
- a no-wifi readiness drill
- a real local-model smoke test through Ollama and OpenClaw inference
- a Codex Mac handoff brief for opening the same workspace in Codex
- optional redacted Markdown/HTML/JSON/SVG audit outputs
That is the point: Clawdeck turns local mode into something you can adopt, verify, and keep using inside your normal setup.
clawdeck adopt [workspace] [--home dir] [--name name] [--force] [--no-agents-link]
clawdeck setup [workspace] [--home dir] [--name name] [--yes] [--smoke] [--timeout ms] [--force] [--no-agents-link]
clawdeck apply [--workspace dir] [--home dir] [--yes]
clawdeck drill
clawdeck smoke [--model ollama/name] [--home dir] [--timeout ms] [--no-openclaw]
clawdeck handoff [--home dir] [--no-checks]
clawdeck audit --out report.md --html report.html --json audit.json --card card.svg
clawdeck local [dir] [--name name] [--force]
clawdeck doctor [--json]
clawdeck snapshot --out setup.json [--home dir]adopt is the main path. It detects your existing OpenClaw workspace from ~/.openclaw/openclaw.json, overlays missing Clawdeck files without overwriting your workspace, and links AGENTS.md to CLAWDECK.md.
setup is the first-run path. It runs adopt, previews or applies the local profile, runs the readiness drill, and leaves one next command. It only writes ~/.openclaw/openclaw.json when you pass --yes.
apply backs up ~/.openclaw/openclaw.json, points the active default model setup at local Ollama models, and preserves existing providers/plugins/auth/gateway settings.
drill checks whether the local path is ready. smoke goes further and asks the selected local model to reply through Ollama and OpenClaw inference.
handoff prints a compact brief you can paste into Codex Mac app so Codex uses the same CLAWDECK.md local-mode contract instead of treating the folder as a random repo.
Clawdeck offline drill: blocked
Local agent stack: primary=ollama/qwen3:4b-instruct, ollamaModels=7, gateway=warn, workspaceFiles=7/7, activeHostedFallback=none.
READY Workspace contract: 7/7 command-center files present
READY Local-model defaults: No hosted model aliases in active defaults
READY Ollama reachable: 7 Ollama models detected
READY Configured model weights: 3/3 configured models installed
READY OpenClaw CLI: doctor=pass
BLOCKED OpenClaw gateway: doctor=warn
Next: openclaw gateway start && openclaw gateway status --json
Clawdeck smoke: pass
Model: ollama/qwen3:4b-instruct
PASS Ollama model reply: ok
PASS OpenClaw local inference: model.run via local provider: ollama model: qwen3:4b-instruct outputs: 1 ok
existing-openclaw-workspace/
AGENTS.md # preserved, with a Clawdeck pointer block
CLAWDECK.md # local-mode switchboard
OFFLINE.md
HEARTBEAT.md
SOUL.md
TOOLS.md
USER.md
.openclaw/
openclaw.template.json
prompts/
launch-brief.md
The active local model lineup:
- primary:
ollama/qwen3:4b-instruct - code:
ollama/qwen2.5-coder:7b - fast fallback:
ollama/llama3.2:3b - hosted fallback: none in the active default model aliases
Clawdeck does not install OpenClaw, install Ollama, pull models, or start daemons for you.
npm install -g openclaw@latest
openclaw onboard --install-daemon
ollama pull qwen3:4b-instruct
ollama pull qwen2.5-coder:7b
ollama pull llama3.2:3b- Clawdeck is MIT-licensed project glue around OpenClaw and Ollama; it is not affiliated with OpenAI, OpenClaw, or Ollama.
- "Free" means no per-token API bill for the local Ollama path. It does not mean free hardware, free electricity, or free optional cloud/search features.
- "Offline" means local files, local shell, local OpenClaw gateway workflows, and already-pulled Ollama models. Network-dependent tasks remain blocked.
- Local small models are useful for many workflows but are not equivalent to latest hosted frontier models for hard reasoning, browser automation, or security-sensitive tool use.
- See docs/claim-check.md for the source-backed claim audit.
Most local-agent setups are either raw dotfiles or separate demo folders. Clawdeck makes local mode a first-class layer in the OpenClaw/Codex workspace you already use: adopt it, prove it, smoke test it, and keep the cloud/local boundary explicit.
Clawdeck does not copy OAuth files, transcripts, browser state, private memory databases, ~/.openclaw/agents, or ~/.openclaw/tasks.
Use clawdeck snapshot when you want to show your setup publicly.
MIT