Built for complex, long-running research
Not another chatbot. MIRA carries "think → experiment → report" end to end and leaves evidence at every step.
Project-level research agent
Built around five primitives — workspace / project / experiment / task_plan — so the agent collaborates long-term rather than chatting once. Every experiment leaves a task_plan.json trace: reproducible and auditable.
20+ LLM providers, out of the box
OpenAI, Anthropic, OpenRouter, DeepSeek, Azure, Ollama, vLLM and more — all built in. Enable routeByComplexity to send cheap jobs to small models and hard jobs to large ones.
CLI / Web / Desktop — three form factors
Run experiments from the command line, watch progress in a browser, or launch the Electron app. They share one config and sync state in real time.
13 channel integrations
Lark, Slack, Telegram, DingTalk, email, Matrix and more. @-mention the agent in a group to run experiments, ship reports, and reply with results — no browser needed.
Self-healing guardrails
Missing fields in an experiment? The model fixes them itself — up to 3 automatic repair rounds before handing off to a human. Strict mode enforces compliant schemas.
Local-first, self-host friendly
All data lives in ~/.mira/. Single-machine, team docker-compose, or remote gateway — three deployment modes, zero vendor lock-in.
A workbench built for agents
MiraUI brings the project queue, pipeline view, experiment detail, and result center together in one Electron app. The agent runs in the background while you flip between projects to check progress — export PPT, papers, or experiment_report with a few clicks.
- Real-time sync: WebSocket push with REST fallback; state comes directly from task_plan.json.
- Same code for Web & Desktop: develop in the browser, ship as Electron, auto-spawns a local mira-engine.
- Pausable & rollback-able: every experiment keeps a full trace of prompts, tool calls, and guardrail fixes.
