OpenClaw + Ollama: the complete setup after Anthropic's third-party clampdown

If you've been on Claude Pro and watched your OpenClaw sessions stall this month, you're in the middle of the largest forced migration the local-AI ecosystem has seen. Google Trends puts openclaw ollama up +507,000% in 30 days. The driver isn't hype — it's Anthropic's April 2026 enforcement against unsanctioned third-party Claude Pro/Max wrappers. OpenClaw and OpenCode users started seeing 429s and capability restrictions unless they upgraded to direct API billing.

OpenClaw was already Ollama-friendly by design. The clampdown just made local the default escape hatch. This guide gets you there.

What you need

Requirement	Minimum	Recommended
OS	macOS 13+ / Ubuntu 22.04+ / Windows 11 (WSL2)	macOS 14+ Apple Silicon, or Linux + NVIDIA
RAM	16 GB	32 GB
GPU VRAM	8 GB (small models)	17 GB+ (27B-class coder models)
Disk	20 GB free	80 GB free (multiple models)
Node.js	20.x	22.x LTS

Apple Silicon with ≥24 GB unified memory handles Qwen 3.6 27B at Q4 comfortably. Intel Macs: skip local, use Gemini CLI free tier instead.

Step 1 — Install Ollama

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

Windows users: install natively from ollama.com/download, not inside WSL. The WSL forwarding works but adds enough latency to ruin the agent loop on smaller models.

Verify and start:

ollama --version
ollama serve &   # local API on http://localhost:11434

Security note. Ollama's default bind is 127.0.0.1. Do not expose 0.0.0.0:11434 to the internet. The Bleeding Llama disclosure in May 2026 confirmed unauthenticated memory disclosure on older versions — upgrade to the latest patched release before exposing anything.

Step 2 — Pull a coding model

For first-time setup, pick one:

Model	Pull command	VRAM (Q4)	Best for
Qwen 3.5 9B	ollama pull qwen3.5:9b	6.5 GB	Entry tier, 16 GB MacBooks
Gemma 4 27B	ollama pull gemma4:27b	17 GB	Balanced quality, 4B active MoE = fast
Qwen 3.6 27B	ollama pull qwen3.6:27b	17 GB	Best coding pick — SWE-bench Verified 77.2, multimodal

Qwen 3.6 27B is the default recommendation. It posts HumanEval 88.5 and SWE-bench Verified 77.2 — beating Qwen 3.5 72B on HumanEval (82.7) while fitting in 17 GB at Q4 — and is explicitly designed to be compatible with Claude Code / Qwen Code tooling, which means OpenClaw's tool-use loop works out of the box. We compare these models in detail in issue #8.

Smoke test:

ollama run qwen3.6:27b "write a python function that reverses a linked list iteratively"

If you can't fit 17 GB of VRAM, see our GPU directory for what each card actually runs.

Step 3 — Install OpenClaw

Check the official OpenClaw repo for the current install command — the project has been renamed twice in the past year. As of 2026-05 the global install is:

pnpm add -g openclaw
# or: npm install -g openclaw

Verify the install:

openclaw --version
openclaw doctor

openclaw doctor checks Node version, Ollama reachability, and write permissions on ~/.openclaw/skills/. Fix anything red before continuing — Skills failing silently later is the most common "OpenClaw doesn't work" report.

Step 4 — Point OpenClaw at Ollama

~/.openclaw/config.json:

{
  "provider": "ollama",
  "endpoint": "http://localhost:11434",
  "model": "qwen3.6:27b",
  "context_window": 131072,
  "temperature": 0.2,
  "skills_dir": "~/.openclaw/skills"
}

Or via CLI:

openclaw config set provider ollama
openclaw config set model qwen3.6:27b
openclaw config set endpoint http://localhost:11434

Start a session:

openclaw

The banner should show provider: ollama and the model name. If it still says provider: anthropic, jump to "Common errors" below.

Step 5 — Install essential Skills

OpenClaw's value comes from its 800+ Skills registry. The minimum useful set for local-model setups:

openclaw skill install code-review
openclaw skill install git-helper
openclaw skill install file-explorer
openclaw skill install web-search       # uses local SearXNG by default
openclaw skill install benchmark        # measures local model latency

Browse the full registry with openclaw skill list --remote. Skill quality varies — code-review, git-helper, and file-explorer are the three that work reliably with any backend model.

Step 6 — First real task

cd ~/your-project
openclaw
> /file-explorer load src/
> refactor the authentication middleware to support OAuth2 PKCE flow

Expect 2–4× slower wall-clock than Claude Sonnet on the same task. No rate limits, no monthly cap, no token bill, code never leaves the machine.

Common errors

Error: connection refused at 127.0.0.1:11434

Ollama isn't running. Start it with ollama serve. Check with lsof -i :11434.

Model not found: qwen3.6:27b

Forgot the pull. Run ollama pull qwen3.6:27b.

Context length exceeded mid-session

Even a 131k window fills fast on big repos. Either switch to a model with a longer effective window via YaRN, or use OpenClaw's /compact Skill to summarize the running session.

OpenClaw still talking to Claude after the switch

Check ~/.openclaw/config.json for a stale anthropic block from a previous openclaw login. Run openclaw logout to clear cached credentials, then re-run the Step 4 config.

Tool calls hang or return empty

Smaller models (≤9B) struggle with multi-step tool use. Either move up to Qwen 3.6 27B, or limit parallelism: openclaw config set max_parallel_tools 1.

What you actually give up

Be honest with yourself before committing:

You lose	You gain
Claude Sonnet-level reasoning on complex refactors	Zero per-token cost
200k cloud context	Sovereignty over your codebase
Best-in-class tool-use reliability	No 429s, no monthly cap
Multimodal image input (on most local models)	Works offline

The pragmatic move: OpenClaw + Ollama for daily edits and refactors, fall back to Claude Code (direct API billing) twice a week for architectural sessions and tricky concurrency bugs. The cost ratio is roughly 1:20 in your favor.

Where to go next

• Full head-to-head of the coding agents you could run on top of Ollama — Claude Code, Cursor, OpenCode, OpenClaw, Gemini CLI, Cluely, z.ai — in issue #8.
• The May 2026 runtime breakthroughs (MTP, DFlash, PAGED MoE) that just made local agentic coding viable: issue #6.
• Pick the right card for the model you want: runlocal GPU directory.