IMA-AI is my personal multi-agent AI platform – a supervisor named IMA that knows you, remembers and delegates tasks to specialised agent teams. Local-first and able to run on your own machine.
Three roles in one interface:
- Personal assistant – IMA organises everyday life & work and delegates to specialist teams.
- AI Workflow Studio – visual automations from building blocks, manually or AI-generated.
- Dev / coding tool – agent teams build software, run code and deploy projects.
How it works
One message becomes a plan: the intent router decides whether IMA answers itself or hands off to a team. A team lead breaks the request into board tasks with dependencies, a board executor runs them as a DAG in parallel, a review pipeline (code + security review) checks each result, and a synthesis pass merges the parts into one coherent result – streamed live back into the chat.
Highlights
- Agentic loop – agents reason, call tools and work in multiple steps (up to 12 iterations).
- Kanban board – autonomous task execution with a review pipeline (planned → in_progress → review → done).
- Workflow engine – visual node editor (~17 node types) with cron scheduling and an AI generator.
- Teams & agents – supervisor → team leads → sub-agents, model configurable per agent.
- Workspace sandbox – sandboxed file system + whitelisted command runner with a live terminal.
- Deploy via Coolify – from project to build & live URL via GitHub push (auto-HTTPS).
- Memory with provenance – facts with provenance and scopes (global/private/work/project).
- Token & cost tracking – usage per agent, daily budget with a fallback model.
Technically a TypeScript monorepo (Turborepo + pnpm): a Fastify 5 backend with built-in node:sqlite (FTS5), a React 19 frontend, LLM access via OpenRouter (plus Gemini and local serving via Ollama). In active development – the core loop works.