Summary

OpenClaw currently connects to Ollama through the OpenAI compatibility layer (/v1/chat/completions), which silently drops tool calls when streaming is enabled. Since OpenClaw hardcodes stream: true, no Ollama model can use tools — the model decides to call a tool, but the streaming response returns empty content with finish_reason: "stop", losing the tool call entirely.

Meanwhile, Ollama's native API (/api/chat) has fully supported streaming + tool calling since May 2025 (blog post, PR ollama/ollama#10415). The problem isn't Ollama — it's that OpenClaw routes through a broken compatibility layer instead of using the native endpoint.

Root causes identified (3 issues)

Proposed solution

Add a dedicated ollama API provider type that talks to Ollama's native /api/chat endpoint directly, with proper streaming chunk handling and context window configuration.

Config example:

{
  "models": {
    "providers": {
      "ollama": {
        "baseUrl": "http://localhost:11434",
        "api": "ollama",
        "models": [{
          "id": "qwen3:32b",
          "name": "Qwen3 32B",
          "reasoning": true,
          "input": ["text"],
          "contextWindow": 131072,
          "maxTokens": 16384
        }]
      }
    }
  }
}

What this enables:

Implementation scope:

1. Add "ollama" to the Api type union
2. Create native Ollama API client (request/response mapping)
3. Handle streaming chunks — accumulate tool_calls from intermediate done:false chunks
4. Set num_ctx from model's contextWindow config (default 65536) to prevent prompt truncation
5. Convert messages/tools between SDK format and Ollama native format

Verified behavior

Tested with qwen3:32b (32B parameters) on MacBook Pro M4 Pro 48GB:

• ✅ curl → Ollama native API with num_ctx=65536 + 23 tools + system prompt → tool_calls generated correctly
• ✅ Streaming text works with all Ollama models
• ✅ Tool call accumulation from intermediate chunks works
• ✅ All 13 unit tests pass

Alternatives considered

• PR fix(ollama): add streamToolCalls fallback for tool calling #5783 (streamToolCalls: false fallback): Disables streaming when tools are present. This works but sacrifices the streaming UX — users see no output until the full response is ready.
• Wait for Ollama to fix /v1/chat/completions: Tracked in ollama#12557, but no timeline. The native API already works, so there's no reason to wait.
• jokelord's supportedParameters patch: Adds config-level tool support declaration for local models (sglang/vLLM). Solves a different problem (tool detection) but doesn't fix the streaming issue with Ollama.

Additional context

• Ollama (and other local models): streaming breaks tool calling — need stream:false fallback #5769 — Original issue: streaming breaks tool calling for Ollama
• fix(ollama): add streamToolCalls fallback for tool calling #5783 — Workaround PR (disable streaming when tools present)
• Ollama streaming tool calling blog post (May 2025)
• ollama/ollama#10415 — Ollama's fix for native API
• ollama/ollama#12557 — OpenAI compat endpoint still broken
• OpenCode issue #1034 — Same num_ctx problem reported in OpenCode project

Tested environment:

• OpenClaw v2026.1.29
• Ollama v0.15.4
• Models: qwen3:32b, glm-4.7-flash, mistral-small3.1:24b, devstral
• OS: macOS (Apple Silicon M4 Pro, 48GB)