Any frontier model. One platform.
Bring your own keys or subscriptions for Anthropic, OpenAI, Google, xAI, Kimi, DeepSeek, Z.ai, or OpenRouter - or point Hezo at Ollama or LM Studio and run on your own hardware. Each model runs inside its own first-party coding CLI, with the same tools, skills, and security underneath.
The meta-harness: a harness around the harnesses.
Hezo doesn't re-implement every provider against a lowest common denominator. Claude runs in Claude Code, GPT in Codex, Gemini in its CLI, Grok in Grok Build, Kimi in Kimi Code, OpenRouter in OpenCode - each model in the harness its maker built for it, wrapped in one uniform platform layer.
- Pick a model per agent - a fast one for triage, a deep one for hard work
- Mix providers freely inside one team
- Live model listing - new provider models appear without an upgrade
A run ends when the work is done, not when the model says so.
When an agent decides it has finished, Hezo judges the work independently before letting the run end. The check is Hezo's, not the model's, so a cheaper model holds the same line as a frontier one.
- Won't let a run stop on failing tests or half-finished work
- Rejects the "out of scope" dodge and the "I'll leave that for later" punt
- Catches a handoff written in the final message that nobody ever received
Budgets with teeth.
Every run is priced from real token counts, tracked per agent, project, and model. Set hard daily, weekly, and monthly caps - at 100% the agent pauses and new work stops. No runaway costs, no surprise bill.
Keep exploring
All features →Bring your models. Cap your spend.
Open localhost:3100 - setup walks you through the rest.