openclaw - 💡(How to fix) Fix [Feature]: Drop-in 5–7× faster tool calls via MTClaw Function Router (100% Pass@1, OpenAI-compatible) [1 comments, 2 participants]

Summary

Document and (optionally) bundle MTClaw Function Router as a first-class OpenAI-compatible provider proxy for OpenClaw, giving users a drop-in 5–7× speedup on tool-heavy workflows.

Problem to solve

OpenClaw users running tool-heavy workloads (desktop control, IDE actions, repeated tool calls) pay the full upstream LLM latency on every turn — even when the routing decision is trivially "call tool X with these args." There is currently no documented path for OpenClaw users to insert a lightweight router in front of their upstream model without writing one themselves. Users either accept the latency, or build their own proxy and miss out on session-aware features (history, deduplication) because OpenClaw's session id is not exposed to provider plugins by default.

Proposed solution

Officially recognize MTClaw Function Router (MooreThreads/MTClaw as a recommended provider proxy for OpenClaw:

1. Docs: Add a short section under OpenClaw's "Providers" / "Integrations" docs page describing how to point base_url at MTClaw's Function Router endpoint. MTClaw is already OpenAI-compatible, so no schema changes needed.
2. (Optional, heavier) Upstream the session-bridge plugin that MTClaw ships (plugins/session-bridge/) so OpenClaw can natively forward its session id to any downstream provider via x-openclaw-session-id. This unlocks accurate session-scoped tool history (/v1/tool_history) for any provider proxy, not just MTClaw.

Demo (the workflow shown is OpenClaw + Function Router vs OpenClaw + raw upstream, side by side):

https://github.com/user-attachments/assets/2ee2a195-fa7f-4be1-a14c-ecdc53bbc83a

Alternatives considered

No response

Impact

• Affected users: OpenClaw users running tool-heavy workloads — desktop automation, IDE-driven coding, evaluation harnesses. Anyone whose upstream model is a frontier-tier general LLM (Doubao, Claude, GPT, etc.) being called for trivial tool routing.
  • Severity: Blocks workflow — current latency makes interactive use feel sluggish; users often abandon agent loops mid-task.
  • Frequency: Always, on every tool turn. Compounds heavily in multi-step tasks (typical agent runs are 5–20+ tool calls).
  • Consequences: Without integration: users either pay 5–7× latency, or have to discover and self-host MTClaw without any guidance from OpenClaw docs. With integration: documented setup path + (optional) native session-id forwarding → faster, more reliable agent runs out of the box.

Evidence/examples

Benchmark on 50 OpenClaw desktop-control tasks, each task run 4 times, baseline = OpenClaw + Doubao (doubao-seed-2-0-pro-260215):

<img width="684" height="198" alt="Image" src="https://github.com/user-attachments/assets/5f7ec31f-2d31-417b-b572-46c23e913658" />

Additional information

No response