A team of AIs,
thinking together.
A guided multi-agent research assistant. Drop in a goal — the planner breaks it into tasks, a research router grounds each one in the web or reasoning, and a writer/critic loop turns it into a cited, polished report. Runs entirely on local LLMs via Ollama.
Four things, done properly.
Every stage is grounded, inspectable, and pauseable — nothing runs as a single opaque call.
Grounded research pipeline
Planner → research router → researchers gathering web or reasoning-based insight, with inline citations, feeding into writer and critic.
Streaming with checkpoints
Server-sent events stream every stage live, with pause/continue and auto-run toggles so you can review before it moves on.
Visual execution graph
A React Flow graph shows planner → router → researchers → the rest of the pipeline, with "needs rerun" badges on stale outputs.
100% local, zero cost
Runs entirely on local LLMs via Ollama — no OpenAI key, no per-call cost, fully offline capable.
How a goal becomes a report.
Goal → planner → research router → researchers (with manual review) → synthesizer → writer → critic.
Plan
You drop in a goal. The planner breaks it into discrete research tasks.
Route & research
Each task is routed to web grounding or LLM reasoning. You can inspect, edit, or rerun researchers before continuing.
Synthesize & write
Cited research blocks are synthesized, then the writer produces a structured, sourced draft.
Critique
The critic reviews and refines recursively until the output holds up.
See it run.
Pick an example goal. Watch it get typed in, then watch the pipeline actually work through it.
Stack
Next.js on the frontend, server-sent events for streaming, and local inference via Ollama — nothing calls out to a paid API.
Clone it, run Ollama, watch it think.
The full source, setup instructions, and architecture notes are on GitHub. No API keys required.