Self-hostable · bring your own key · runs on local models
Multi-Agent Research Assistant
Cited research you can actually verify.
A multi-agent pipeline that searches, gathers evidence, and writes a cited report — then pauses for you twice, and exports something a third party can check offline, with no AI and no network.
New in v3.1.0Research only the files you choose, with every claim checked against them before you see the report. What changed →
What this does that others do not
Every citation is falsifiable
Each [n] resolves to a real source and the verbatim sentence that supports it. One that cannot be verified renders a ⚠ chip instead of rendering clean.
You approve the plan before it spends
The run pauses after the planner and shows the research areas it chose. Drop what you did not ask for and it is never researched — all before a single search costs anything.
Your corpus never has to leave
Corpus-only mode takes evidence only from your documents, with web search off and indexing on a local model. The research models still see your question and excerpts, so choose local models too and an unpublished manuscript stays on your machine.
What it does
Everything here ships in v3.1.0. Where a feature has a limit that decides whether it applies to you, its description says so.
Your instructions for each agent
Rewrite how the planner, executor, critic, synthesizer and follow-up chat behave. The checks that keep research honest cannot be rewritten, and a run’s verification bundle records exactly which instructions produced it.
Read more →A record, not just a report
Evidence, sources, claims, the evidence each claim resolved to, and conflicting sources are records you can inspect. A claim that resolved to no evidence says so instead of looking supported.
Read more →Any model, for each agent
Route each agent to Google, Anthropic, OpenAI, OpenRouter, any OpenAI-compatible endpoint, or a local model through Ollama — on your own keys, with nothing proxied through a service of ours.
Read more →Chat over approved research
Project chat answers from the reports you approved in a project and cites the ones it used, so an answer traces back to the research behind it. Drafts and rejected work never reach it. Self-hosted server only.
Read more →Desktop app or your own server
Install on an Apple Silicon Mac (Windows and Linux: the last build published for them) with no Docker, database or login to set up, or self-host the full stack with one command. Desktop builds are unsigned, so macOS and Windows warn on first launch.
Read more →Graded by an independent judge
Citation support measured 96.4% on ten fixed questions, graded by a model from a different vendor than every model under test. One run on one model routing — and published as exactly that.
Read more →
Two human gates, not zero
Both are durable checkpoints, not polling loops: state is written to disk and the worker exits, so resuming continues rather than re-running research you already paid for.
- 01Plannerdecomposes the question
- 02⏸ Design gateyou choose what gets researched
- 03Executor ⇄ Criticgathers and grades evidence
- 04Synthesizerwrites the cited draft
- 05⏸ Review gateyou approve the draft
- 06Artifactfrozen when you approve, verifiable offline
The part nobody else has
Every report exports as a .bundle.json carrying its evidence, its sources, the models actually dialled, the instructions each agent ran with, and the approval chain — hashed so that editing the report after a human approved it breaks the chain. A standalone verifier checks the whole thing with no AI, no network, and no account. Your reviewer does not have to trust this tool, or us.
Open source, and inspectable
MIT licensed. The verification claims on this site are the kind that are worth nothing unless you can check them, so the source, the tests that hold them up, and the documentation are all public and all built from the same repository as this page.
- Licence
- MIT. Use it, fork it, run it commercially — no per-seat cost and no lock-in.
- Your keys, your data
- Bring your own API key, or run entirely on local models. Nothing is proxied through a service we operate, because there isn't one.
- Auditable claims
- The measurements this project publishes are backed by committed evaluation results, and the harness refuses to print a number it did not measure.