OpenClaw

Verified against docs.openclaw.ai on 2026-10-11 (sidebar release v2026.9.9; sources in research/clients-2026-10-11.md, §2). OpenClaw is the project formerly released as Warelay → Clawdbot → Moltbot (renamed OpenClaw 2026-01-30). Test status: configuration keys documented; live OpenClaw tool round trip NOT RUN in this build.

Generate the snippet: modelsaar client-config openclaw (--apply merges into ~/.openclaw/openclaw.json with a backup when the file is plain JSON; JSON5 files with comments must be merged by hand).

Key facts

• Config file ~/.openclaw/openclaw.json (JSON5). Provider block models.providers.<id> with baseUrl (include /v1), apiKey (${ENV_VAR} substitution), api: "openai-completions" (Chat Completions — the default when omitted; do not use openai-responses, ModelSaar returns 501 for /v1/responses).
• Custom provider models are not auto-discovered: declare each alias under models[] with id, name, reasoning, input (["text"] or ["text","image"]), contextWindow, maxTokens, cost {input, output, cacheRead, cacheWrite} (zeros for local).
• compat.supportsPromptCacheKey: true makes OpenClaw send prompt_cache_key (session id + boundary count) to a custom endpoint; ModelSaar reads it (affinity.body_fields) to keep a run on one deployment. compat.supportsTools should mirror the alias contract.
• Default model: agents.defaults.model.primary: "modelsaar/auto"; aliases via agents.defaults.models."modelsaar/auto".alias.
• request.allowPrivateNetwork defaults to false but a custom baseUrl origin is trusted exactly.

{
  agents: { defaults: { model: { primary: "modelsaar/auto" } } },
  models: {
    mode: "merge",
    providers: {
      modelsaar: {
        baseUrl: "http://127.0.0.1:8484/v1",
        apiKey: "${MODELSAAR_API_KEY}",
        api: "openai-completions",
        timeoutSeconds: 300,
        models: [
          { id: "auto", name: "ModelSaar auto", reasoning: false, input: ["text"],
            contextWindow: 32768, maxTokens: 8192,
            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
            compat: { supportsTools: true, supportsPromptCacheKey: true } },
          { id: "auto-code", name: "ModelSaar auto-code", reasoning: false, input: ["text"],
            contextWindow: 32768, maxTokens: 8192,
            cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
            compat: { supportsTools: true, supportsPromptCacheKey: true } }
        ]
      }
    }
  }
}

export MODELSAAR_API_KEY=$(head -1 ~/.modelsaar/keys/api_keys) in the shell that runs OpenClaw.

Use the contextWindow printed by client-config (the alias contract's minimum across members) so OpenClaw's own context management stays within what every eligible deployment can serve.

Notes

• OpenClaw streams and uses a 300 s first-response budget for local endpoints; timeoutSeconds above matches it.
• utilityModel (OpenClaw's cheap internal-task model) can point at modelsaar/auto-fast.
• Whether OpenClaw sends tool_choice/parallel_tool_calls is not documented; if it does send tool_choice: required, Ollama-backed deployments are ineligible and ModelSaar returns 422 with that blocker.