v2.4.0: stop treating Claude like a chat tab
openclaw-claude-proxy is not trying to imitate Claude's native UI.
It is the stable local API bridge for the rest of your tool stack.
If Claude Code channel is the native in-app experience, OCP is the bridge that makes Claude usable from everything else: OpenClaw, Cursor, Continue, Open WebUI, LangChain, custom routers, and any OpenAI-compatible client.
What's actually better in v2
No more warm-pool fragility
- v2 uses on-demand spawning
- no pre-warmed worker pool
- no stale workers
- no degraded warm state to babysit
Real-world local workflow performance
- real streaming instead of buffered upstream responses
- adaptive timeout tiers by model class
- faster and cleaner failover behavior
- in practice, local tool / agent workflows are now competitive with Claude channel latency, while remaining easier to route and diagnose
More stable where channel-style workflows get brittle
- better fit for headless automation
- better fit for OpenAI-compatible tools that know nothing about Claude channel
- explicit
/health,/sessions, diagnostics, auth state, and local observability
Security hardening
- stricter request validation
- stronger sanitization on the proxy boundary
- safer behavior under malformed input / timeout edges
Included in this release
- Per-model circuit breaker — one model failing no longer poisons the whole proxy
- Adaptive timeouts — Opus and other heavier models get saner timeout budgets
- Real streaming — upstream output is actually streamed
- Security hardening — tightened request handling and validation
Bottom line
Claude channel is great inside Claude Code. OCP is better as the stable local API bridge for everything else.
Full Changelog: v2.3.0...v2.4.0