Manifest Weekly #05

Big week. Manifest now works with several AI agent, not just OpenClaw. We also shipped routing by task type and moved self-hosting to Docker-only.

What we shipped this week

Manifest works with any AI agent Manifest is no longer tied to OpenClaw. Any agent that speaks OpenAI-compatible HTTP can route through it: Hermes, Vercel AI SDK, LangChain, OpenAI SDK, or anything you build with cURL. The setup flow now lets you pick your agent framework and gives you the right instructions for it.

Routing by task type On top of the existing complexity tiers (simple/standard/complex/reasoning), you can now assign specific models to specific task types. Coding, web browsing, trading… Each gets its own primary and fallback model. If your agent mostly writes code, you can point coding tasks at a strong coding model and let everything else route cheaper.

Ollama Cloud and Z.ai GLM Coding Plan Two new subscription providers. Ollama Cloud connects with an API key from ollama.com. Z.ai gives access to GLM models through their token-based coding plan.

Docker-only self-hosting The OpenClaw plugin is gone. Self-hosting is Docker-only now: pull the image, run it, point your agent at it. No more plugin compatibility issues across OpenClaw versions. The image runs on both amd64 and arm64.

Fixes

• Gemini parallel tool calls and Anthropic extended thinking were losing state during fallback round-trips (#1517, #1519)
• Short greetings with tools attached no longer get routed to complex tier (#1560)

What’s coming next

• Custom header-triggered routing tiers
• One API key for all providers

— The Manifest team