The best LLM for OpenClaw
OpenClaw is an autonomous agent — it runs long chains of tool calls against your files, shell and apps. That puts unusual demands on the model behind it: one task can fire dozens of calls, and the moment a single provider rate-limits mid-loop or a model returns an empty tool response, your agent stalls. CodeBurst gives OpenClaw a model endpoint built for exactly this: self-healing failover across many providers, and tool-call repair, behind one model name.
What an agent needs from its model backend
A chatbot makes one call per message. An agent like OpenClaw makes many — read a file, decide, run a command, read the result, decide again. Two failure modes that barely matter for a chatbot become loop-killers for an agent:
- Rate limits mid-task. Burst traffic from a single agent run trips a provider's per-minute cap halfway through a job. The agent doesn't get a graceful "try later" — it gets an error in the middle of a multi-step task.
- Empty tool responses. Several strong reasoning models return empty content on the turn where they're supposed to synthesize a tool result back into an answer. The agent receives nothing and stalls — even though the model is perfectly capable.
OpenClaw has its own model-failover list, but it fails over between models you enumerate. The gap underneath is provider reliability and tool-format handling — and that's what a router solves.
How CodeBurst keeps the loop running
| Agent failure | What CodeBurst does |
|---|---|
| Provider rate-limits mid-loop | Every model name is a multi-provider chain — CodeBurst reroutes to the next healthy provider in the same request, so the agent never sees the outage. |
| Reasoning model returns empty on a tool turn | codeburst-agent detects the empty tool-synthesis response and retries with a corrected format, so the agent gets a real answer instead of a stall. |
| Agent context overflows the window | Route a compaction step through codeburst-compress — size-aware, so it never sends more than a model can hold. |
| Agent needs to read a screenshot | codeburst-vision handles images with its own vision-model failover chain. |
Point OpenClaw at CodeBurst
CodeBurst is OpenAI-compatible, so it slots in as a custom provider. In your OpenClaw models.json:
{
"models": {
"providers": {
"codeburst": {
"baseUrl": "https://codeburst.ai/api/v1",
"apiKey": "YOUR_CODEBURST_KEY",
"models": {
"codeburst-agent": {},
"codeburst-best": {},
"codeburst-swarm": {}
}
}
}
},
"agents": {
"defaults": {
"model": {
"primary": "codeburst/codeburst-agent",
"fallbacks": ["codeburst/codeburst-best", "codeburst/codeburst-swarm"]
}
}
}
}Now OpenClaw's own fallback list sits on top of CodeBurst's per-provider failover — two layers of resilience instead of one. Each entry like codeburst-agent is already a self-healing chain, so even your primary never depends on a single provider being up.
Which alias for which job
codeburst-agent— your default. Tuned for tool-calling, with tool-format repair built in.codeburst-swarm— hard reasoning steps; multiple models vote for a more reliable answer.codeburst-vision— when the agent needs to read a screenshot or photo.codeburst-compress— compact a long agent context without overflowing the model's window.codeburst-best— "just give me the strongest model that's up right now."
Get started
Grab a key, drop the provider block into models.json, and run OpenClaw as usual — now on an endpoint that routes around provider failures for you.
FAQ
Can CodeBurst be OpenClaw's model backend?
Yes — add it as a custom OpenAI-compatible provider (baseUrl https://codeburst.ai/api/v1) and point the agent at codeburst-agent.
Why does my agent stall on one provider?
Long tool loops trip rate limits mid-task, and some reasoning models return empty tool responses. CodeBurst fails over across providers and repairs empty tool turns so the loop keeps moving.
Is it OpenAI-compatible?
Yes. Standard /v1/chat/completions — change the base URL and key and OpenClaw works.