How to Fix the "ollama launch openclaw" Error: unknown integration

You installed Ollama, pulled a model, and ran the one command every tutorial mentions: ollama launch openclaw. Instead of a friendly setup wizard, the terminal spat back Error: unknown integration. Nothing launched, nothing was configured, and the error gives you exactly zero hints about what to do next. The good news: this is one of the easiest OpenClaw errors to fix, and there is a reliable manual path that works even if the launcher never cooperates.

Error: unknown integration

This error comes from the Ollama CLI, not from OpenClaw. It means the Ollama binary you invoked does not have openclaw in its table of launchable integrations. The ollama launch openclaw onboarding flow shipped in Ollama v0.17.0 (February 2026). Any older build has never heard of the subcommand target, so it fails with "unknown integration" before OpenClaw is ever contacted.

Three causes account for almost every report:

• Outdated Ollama. You are running a build older than v0.17.0.
• A stale binary on your PATH. You updated Ollama, but an older copy (a leftover Homebrew install, or an old manual install) sits earlier in your PATH and gets picked up first.
• A typo or missing OpenClaw install. The integration name must be exactly openclaw, and OpenClaw itself must be installed for the handoff to complete.

Fix it step by step

1. Check your Ollama version.

   ollama --version

   If it prints anything below 0.17.0, that is your problem. Skip to step 2. If the version looks current, jump to step 3.
2. Update Ollama. Download the latest build from ollama.com/download (or update through your package manager), then restart the Ollama app or daemon so the new binary is actually the one serving requests.
3. Check which binary you are really running.

   # macOS / Linux
   which ollama
   
   # Windows PowerShell
   where.exe ollama

   If the path points somewhere unexpected (an old install location), remove the stale copy or reorder your PATH so the fresh install wins.
4. Confirm OpenClaw is installed.

   openclaw --version

   If that command is not found, install OpenClaw first; the launcher cannot hand off to software that is not there.
5. Retry the launch.

   ollama launch openclaw

   On v0.17.0 or newer with a clean PATH, the integration handshake should now start.

The reliable fallback: configure the provider manually

The launcher is only a convenience wrapper. Everything it does can be done directly in OpenClaw, and the manual route works on any Ollama version. It is also the better option when Ollama runs on a different machine than your Gateway.

1. Pull a model and note its exact name.

   ollama pull llama3.3
   ollama list

2. Set the local credential marker. Local and LAN Ollama hosts do not need a real key; OpenClaw uses a marker value:

   export OLLAMA_API_KEY="ollama-local"

3. Run OpenClaw onboarding and pick Ollama.

   openclaw onboard

   Select Ollama, then a mode (Local only is the usual choice). OpenClaw discovers your installed models automatically. Prefer config files? Add the provider explicitly:

   {
     models: {
       providers: {
         ollama: {
           baseUrl: "http://127.0.0.1:11434",
           apiKey: "ollama-local",
           api: "ollama",
           models: [
             { id: "llama3.3", name: "llama3.3", input: ["text"] }
           ]
         }
       }
     }
   }

4. Select the model.

   openclaw models set ollama/llama3.3

5. Verify.

   openclaw models list --provider ollama

Do not add /v1 to the base URL

OpenClaw talks to Ollama's native API at http://host:11434, not the OpenAI-compatible /v1 endpoint. Adding /v1 silently switches to compatibility mode, where tool calling breaks and models can emit raw tool-call JSON as plain text. If your agent replies look like machine gibberish, check the base URL first.

Model names must match exactly

Model ids in OpenClaw must be exactly what ollama list prints, including tags. If Ollama shows llama3.3:latest, referencing llama-3.3 or a guessed name fails as an unknown model. Copy and paste, do not retype.

Still failing? Two quick diagnostics

First, confirm the Ollama daemon is actually reachable from the machine running the Gateway:

curl http://127.0.0.1:11434/api/tags

If curl works but OpenClaw does not, your Gateway probably runs in Docker or on another host where localhost points somewhere else; use the LAN address in baseUrl instead. Second, run a minimal smoke test that skips the full agent stack:

openclaw infer model run --local \
  --model ollama/llama3.3 \
  --prompt "Reply with exactly: pong"

If the probe succeeds while normal agent replies fail, the endpoint is fine and the issue is the model's capacity for agent tool schemas; try a larger model or a leaner agent profile.

Do I need a real API key to use Ollama locally with OpenClaw?

No. Loopback, private-network, and .local Ollama hosts only need the marker value ollama-local as the key. A real credential is required only for public remote hosts or the hosted ollama.com service.

Which Ollama version do I need for ollama launch openclaw?

Ollama v0.17.0 or newer, since that release added OpenClaw to the launch integration table. Older versions always print "unknown integration". You can still connect older Ollama builds to OpenClaw by configuring the provider manually.

Can OpenClaw use Ollama running on a different machine?

Yes. Point baseUrl at the remote host, for example http://gpu-box.local:11434, keep api: "ollama", and list the models manually since auto-discovery is disabled for non-loopback custom hosts. Make sure the remote daemon binds to the LAN interface and any firewall allows port 11434.