01 / agents
Grounded research pipeline
Planner -> research router -> researchers gathering web or reasoning-based insight, with inline citations, feeding into writer and critic.
100% local - zero API cost
A team of AIs,
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.
execution_graph.live
runtime status
Inspectable, checkpointed, and built to stay local.
Every stage can be paused, reviewed, and rerun before the final report lands.
Streaming checkpoints
pause / continue
Execution graph
planner -> router -> writer
Grounded output
inline citations kept
Local inference
Ollama only
// what it does
Four things, done properly.
Every stage is grounded, inspectable, and pauseable - nothing runs as a single opaque call.
02 / control
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.
03 / graph
Visual execution graph
A React Flow graph shows planner -> router -> researchers -> the rest of the pipeline, with "needs rerun" badges on stale outputs.
04 / cost
100% local, zero cost
Runs entirely on local LLMs via Ollama - no OpenAI key, no per-call cost, fully offline capable.
// request lifecycle
How a goal becomes a report.
Goal -> planner -> research router -> researchers (with manual review) -> synthesizer -> writer -> critic.
01
Plan
You drop in a goal. The planner breaks it into discrete research tasks.
02
Route & research
Each task is routed to web grounding or LLM reasoning. You can inspect, edit, or rerun researchers before continuing.
03
Synthesize & write
Cited research blocks are synthesized, then the writer produces a structured, sourced draft.
04
Critique
The critic reviews and refines recursively until the output holds up.
// try it
See it run.
Pick an example goal. Watch it get typed in, then watch the pipeline actually work through it.
goal.txt
process.log
// built with
Stack
Next.js on the frontend, server-sent events for streaming, and local inference via Ollama - nothing calls out to a paid API.
TypeScript
Next.js (App Router)
Material UI
React Flow
Server-Sent Events
Ollama (local LLM)
Clone it, run Ollama, watch it think.
The full source, setup instructions, and architecture notes are on GitHub. No API keys required.