Real-time galaxy
Watch the graph grow live over SSE as your agent works — then replay the whole run step by step, with every claim's score as it stood at that moment.
A monitor for autonomous research agents
Watch your research agent earn its conclusions.
Auto-Research Monitor watches an autonomous research agent while it works. Every question, experiment, run, result and claim lands on a live evidence lineage graph the moment it happens, each claim scored green, yellow or red by how strong the evidence beneath it really is — and Replay shows exactly how the agent got there.
task/credit-risk · live
conclusion
claim
strong
mixed
weak
Early access
Ask for the code, then run it in two commands.
The source isn't public yet. Email to request early access, or try the live demo in your browser meanwhile. Once you have the code there is no build step and no frontend toolchain.
# 1 — install dependencies
$ pip install -r requirements.txt
# 2 — launch the sidecar server
$ python run.py
→ Open http://localhost:8000
What it does
Make an agent's reasoning legible — and checkable.
Six capabilities turn a stream of agent events into a scored, traceable map of how every conclusion was earned.
Sidecar for any agent
Claude Code, GPT, or your own scripts push events over a simple HTTP API or the zero-dependency Python SDK. Humans watch at a shareable URL.
Credibility scoring
Every claim is scored across 5 structural dimensions — evidence coverage, reproducibility, data provenance, statistical validity and pipeline integrity — then rated green, yellow or red.
Lineage tracing
Click any claim to highlight its full evidence chain back to the original motivation — or to reveal that it has none.
File-watcher inference
Point it at a project folder and files become typed nodes automatically — code, data, runs, results, claims — with edges inferred from temporal proximity.
Convergence & conclusions
Collapse dozens of results into ≤5 apex findings and generate a single traceable conclusion report — heuristic, or written by an LLM.
How it works
From a stream of events to an earned conclusion.
An append-only pipeline: nothing is overwritten, so the whole chain of reasoning stays auditable.
01
Your agent pushes events
It POSTs create_node / create_edge to the events API — or simply writes files into a watched folder.
02
The server logs the graph
FastAPI records an append-only evidence graph in NetworkX — typed nodes, typed edges, full history.
03
The browser renders a galaxy
Cytoscape.js streams updates over SSE into a live, scored galaxy that anyone can watch at /t/{id}.
04
Converge into a conclusion
Results collapse into ≤5 apex findings and a single traceable conclusion report — heuristic or LLM-written.
Five ways to see it
Watch it map real reasoning, live.
Five bundled demos, from a streaming SDK build to a real Claude Opus agent being graded on its own evidence.
⧉ Open the live demoFour finished example runs, read-only in your browser. The five demos below come with the app, which is in early access.
Demo 01
Live SDK build streaming
A Credit-Risk evidence graph streams in node-by-node through the Python SDK, so you can watch the galaxy assemble itself in real time.
live SSE node-by-node
Demo 02
ML sentiment static graph
A finished sentiment-classification project rendered as a static graph — a clear green / yellow / red spread across its claims.
greenyellowred
Demo 03
Finance file-watcher inferred
Files are written to disk on a timer; the watcher types each one and infers edges live from temporal proximity — no manual wiring.
auto-typed watchdog
Demo 04
Auto-research fact-checking verified
Claims are checked against re-runnable sources — including a real GPU data-leakage experiment where a "ship it at 100%" claim is flagged red.
100% → red flag
Demo 05 · featured
crisis-pred real Opus 4.8 agent
A real Claude Opus 4.8 agent researches "Is a systemic financial crisis likely in 2026?" while a cartographer watches its files and maps an earned credibility graph — every claim scored only by the evidence the agent actually produced.
3 green4 yellow3 red
How a score is earned
Five structural dimensions. One honest color.
Scores measure the structure of the evidence chain — not how confident the prose sounds.
Evidence Coverage Is the claim backed by results, runs and artifacts?
30%
Statistical Validity Are the results statistically sound?
20%
Reproducibility Can the supporting runs be re-executed?
20%
Data Provenance Is the data traceable to a real source?
15%
Pipeline Integrity Is the chain from data to claim unbroken?
15%
GreenEarned — strong, traceable evidence.
≥ 72%YellowPartial — mixed or incomplete support.
≥ 40%RedWeak — little to no real evidence chain.
< 40%