Claw Code – Offline / Ollama Mode

Run the full Claw Code agent 100 % locally – no api key, no cloud, no internet required.

---

What changed from upstream

Area	Change
rust/crates/api/src/providers/mod.rs	Added ProviderKind::Ollama, Ollama model registry entries, ollama_is_available() probe, auto-detection logic
rust/crates/api/src/providers/openai_compat.rs	Added DEFAULT_OLLAMA_BASE_URL, OLLAMA_DUMMY_API_KEY, OpenAiCompatConfig::ollama()
skills/connectivity.py	Online/offline detection + DuckDuckGo web search fallback
skills/rag_skill.py	Local ChromaDB RAG over .txt/.md knowledge files
skills/tts_skill.py	Offline text-to-speech via pyttsx3 (espeak backend)
skills/sandbox_skill.py	Isolated code execution via docker run --rm
ollama-config.json	Example config to drop into ~/.claw.json

---

1 – Prerequisites

# Rust toolchain
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

# Python 3.10+
python --version

# Docker (for sandbox skill)
# https://docs.docker.com/engine/install/

# Ollama
curl -fsSL https://ollama.com/install.sh | sh

---

2 – Pull local models

# Primary coding model (~400 MB, runs on CPU)
ollama pull qwen2.5-coder:0.5b

# General-purpose (recommended upgrade)
ollama pull gemma2:2b

# Ultra-fast tiny model for simple tasks
ollama pull smollm:135m

Verify Ollama is running:

ollama list
curl http://localhost:11434/v1/models   # should return JSON

---

3 – Configure Claw Code

Copy the bundled config to your home directory:

cp ollama-config.json ~/.claw.json

Or set environment variables instead:

export OLLAMA_BASE_URL="http://localhost:11434/v1"
export OLLAMA_API_KEY="ollama"   # any non-empty string works

To use the coding model for all requests:

# in ~/.claw.json
{ "model": "qwen2.5-coder:0.5b" }

# or pass on the CLI
claw --model qwen2.5-coder:0.5b "refactor this file"

---

4 – Build Claw Code (Rust)

cd rust
cargo build --release
# binary at: rust/target/release/claw

---

5 – Install Python skills

# Core (required for all skills)
pip install pyttsx3

# RAG skill
pip install chromadb sentence-transformers

# Web search fallback (online mode)
pip install duckduckgo-search

# Linux TTS backend
sudo apt install espeak espeak-ng libespeak-dev   # Debian/Ubuntu

---

6 – Using the Skills from Python

RAG – index your knowledge base

from skills.rag_skill import RagSkill

# Put .txt / .md files in ./knowledge/
rag = RagSkill(knowledge_dir="./knowledge")
rag.ingest()                              # index all files (idempotent)

results = rag.search("how to handle async errors in Python")
for r in results:
    print(f"[{r['score']}] {r['source']}\n{r['text']}\n")

Connectivity – auto web-search or RAG

from skills.connectivity import search_or_rag
from skills.rag_skill import RagSkill

rag = RagSkill()
rag.ingest()

# Automatically uses web search when online, RAG when offline
hits = search_or_rag("Python memory leak debugging", rag_fn=rag.search)

TTS – speak agent output

from skills.tts_skill import TtsSkill

tts = TtsSkill(rate=160)
tts.speak("Build succeeded. 0 errors, 2 warnings.")

# List available voices
for v in tts.list_voices():
    print(v)

# Save to file
tts.save("Task complete.", path="done.wav")

Sandbox – run untrusted code safely

from skills.sandbox_skill import SandboxSkill

sb = SandboxSkill(network="none", memory="128m", timeout=15)

# Python
r = sb.run_python("import math; print(math.pi)")
print(r.stdout)         # 3.141592653589793
print(r.success)        # True

# Bash
r = sb.run_bash("echo $((2**10))")
print(r.stdout)         # 1024

# Node.js
r = sb.run_code("node -e 'console.log(process.version)'", image="node:20-alpine")

---

7 – Model selection guide

Task	Recommended model	Flag
Code editing / generation	qwen2.5-coder:0.5b	default
General Q&A	gemma2:2b	--model gemma2:2b
Ultra-fast completions	smollm:135m	--model smollm:135m
Larger context / quality	qwen2.5-coder:1.5b	--model qwen2.5-coder:1.5b

---

8 – Offline vs Online behaviour

Internet available?
       │
       ├─ YES → web_search via DuckDuckGo (no key needed)
       │          + full Claw Code tool use
       │
       └─ NO  → local RAG over ./knowledge/
                  + TTS announcements
                  + Sandbox execution
                  (all core agent tools still work)

The connectivity check (skills/connectivity.py) is a lightweight TCP probe to 1.1.1.1:53 – it adds < 2 ms latency and causes no network traffic when offline.

---

9 – Troubleshooting

Ollama not detected

ollama serve          # start the daemon if it isn't running
curl localhost:11434  # should print Ollama version

"no voices found" (TTS)

sudo apt install espeak-ng   # install system TTS engine
python -c "import pyttsx3; e=pyttsx3.init(); print(e.getProperty('voices'))"

Docker sandbox errors

docker info           # check daemon is running
docker pull python:3.12-slim   # pre-pull default image

ChromaDB import error

pip install --upgrade chromadb sentence-transformers

---

Licence

Same as upstream Claw Code (see LICENSE).
Skill additions © 2026 – MIT.