openclaw - 💡(How to fix) Fix [Bug]: Ollama remote host autodiscovery impossible — schema validation contradicts documentation [1 participants]

Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

The documentation states autodiscovery activates when models.providers.ollama is not defined, but for remote Ollama hosts the schema validator rejects any config without a provider block, making autodiscovery structurally impossible for remote setups.

Steps to reproduce

1. Configure remote Ollama at a non-localhost address
2. Attempt Option A — remove models.providers.ollama block entirely: Result: schema validates OK but discovery targets hardcoded 127.0.0.1:11434, cannot reach remote host
3. Attempt Option B — keep provider block, remove models array: Result: schema validation FAILS: "models.providers.ollama.models: Invalid input: expected array, received undefined"
4. Attempt Option C — keep provider block with models: []: Result: schema validates BUT autodiscovery disabled because explicit provider block exists
5. Set plugins.entries.ollama.config.discovery.enabled: true: Result: no effect when explicit provider block is present

Expected behavior

Per docs.openclaw.ai/providers/ollama (Custom base URL tab): a provider block with baseUrl and no models array should be valid config. Per the discovery section: autodiscovery should activate when no explicit models are defined. Remote host users should be able to get autodiscovery by specifying baseUrl without a models array.

Actual behavior

No configuration combination enables autodiscovery from a remote Ollama host:

• Without provider block: discovery hardcoded to 127.0.0.1 (unreachable)
• With provider block + no models array: schema validation fails
• With provider block + models: []: discovery disabled by explicit provider block

openclaw models list --local always returns "No models found" regardless of configuration attempted.

Only working approach: manually enumerate all models in models array, defeating the purpose of autodiscovery.

Notable: When the OpenClaw agent uses its exec tool to directly query Ollama: curl http://192.168.40.22:11434/api/tags It successfully returns all 22 models. The remote host is reachable and the models are available — the failure is specifically in OpenClaw's autodiscovery mechanism, not in network connectivity or Ollama availability.

OpenClaw version

2026.4.22 (also reproduced on 2026.4.12)

Operating system

Ubuntu 24.04 (running as Unraid VM)

Install method

npm global

Model

ollama/qwen3.5:27b

Provider / routing chain

openclaw -> ollama (remote host 192.168.40.22:11434)

Additional provider/model setup details

Remote Ollama at http://192.168.40.22:11434 — confirmed reachable: curl http://192.168.40.22:11434/api/tags returns all 22 models correctly

openclaw models list --local: "No models found" (confirmed bug) openclaw models list: shows only models explicitly listed in agents.defaults.models plugins.entries.ollama.config.discovery.enabled: true — has no effect

Logs, screenshots, and evidence

Impact and severity

Affected: All users running Ollama on a separate host from OpenClaw Severity: Medium — workaround exists (manual model enumeration) Frequency: 100% reproducible for any remote Ollama setup Consequence: Users must manually maintain model list in config; new models pulled to Ollama are not visible in OpenClaw UI until manually added to openclaw.json

Additional information

Suggested fix options:

1. Allow empty models: [] to trigger discovery against configured baseUrl instead of disabling discovery entirely
2. Add plugins.entries.ollama.config.discovery.baseUrl override that takes precedence over hardcoded 127.0.0.1 default
3. Update docs to accurately reflect that autodiscovery is not supported for remote Ollama hosts (if this is intentional behavior)

22 models confirmed available on remote Ollama via /api/tags — none are visible to OpenClaw without manual enumeration.

The OpenClaw agent itself can enumerate all 22 remote Ollama models via direct API call (curl /api/tags), confirming the remote host is fully reachable from the OpenClaw VM. This makes the autodiscovery failure purely a framework-level issue, not a network or Ollama configuration issue.