Overview

A self-hostable, bring-your-own-key research assistant with two human checkpoints and citations you can verify.

You ask a research question. A pipeline of specialised agents plans the work, searches, gathers evidence with sources, grades that evidence, and drafts a cited Markdown report. It pauses twice for you — once on the plan, once on the draft — and the finished report exports to a file a third party can check offline with no AI and no network.

The problem it solves

Every general-purpose assistant will write you a research report. The difficulty is not getting one; it is knowing which sentences in it are true.

A citation in most tools is a link. It tells you the model saw a page, not that the page says what the sentence claims. When independent testing looks at this closely, misattribution is common enough that a reader has to re-check the work by hand — at which point the report saved them nothing.

This project treats a citation as a data structure instead of a formatting convention:

• every [n] resolves to a source and the verbatim snippet that supports that claim;
• a marker that cannot be resolved renders a visible ⚠ unverified chip rather than rendering clean;
• the .bundle.json export carries the claims, the snippets, their content hashes, and the approval trail, and ships with a standalone verifier that needs no network.

The system is built to show its own failures rather than to look finished.

Who it is for

What makes it different

Three properties, and it is the combination that is unusual rather than any one of them.

1. The human gate is a durable checkpoint, not a status flag. Both pauses are LangGraph interrupt() calls. The graph checkpoints to the database and the worker process exits. Approving hours later resumes the same state — it does not re-run the research you already paid for. A regression test asserts the planner is invoked exactly once across submit and approve.

2. You approve the plan before it spends anything. The first gate sits between the planner and the first search. Edit the subtopics, pick the report structure, drop a task you did not ask for — and the dropped task is never researched. That is the difference between "the agent picked six queries" and "these are my six subtopics, in my structure".

3. Verification is measured, and unmeasured is not zero. citation_resolution_rate is recorded for every approved report and shown on every row of history — including a Not measured band, because "made no citable claims" and "every marker is broken" are opposite findings. The same rule runs through the evaluation harness: it returns None where it could not measure, and refuses to print a number it did not take. Its citation-support grade comes from a judge that is none of the models under test, and the result names the model that actually answered each ruling.

What it does today

Details: Running research · Citations · Projects and memory · Agent instructions

What is deliberately out of scope

Saying no is what keeps the parts above honest.

• Billing, plans, or payments. Bring your own key; the operator pays for what they use.
• Teams, organisations, roles, shared workspaces. Single-tenant by design.
• Social login / OAuth providers. Email and password, so a self-host needs no external identity provider.
• Being a general chat assistant. Chat here interrogates verified research. It is not a research tool that happens to chat.
• Kubernetes manifests, autoscaling, sharding. Not until a real scaling need exists; the operations page states the order those needs would actually arrive in.
• Mobile apps.

Known limits, stated plainly

• Retrieval quality is the ceiling. With no Tavily or Brave key the chain falls back to DuckDuckGo, which is rate-limited and slow. Report quality tracks retrieval quality more than it tracks model quality.
• Evidence de-duplication is URL-exact. The same article at two URLs becomes two sources.
• The critic grades per-task evidence, not the finished report. Whether the synthesizer's [n] usage is faithful to the snippet it points at is measured offline by the evaluation harness, not by a runtime gate.
• Cost caps cannot fire on OpenRouter or custom endpoints, whose prices are not in the catalog. Cap spend at the provider. See Configuration.
• Small local models fail the structured-evidence step. 14B is the working floor for research; anything runs chat. See Local LLM setup.
• Citation support rests on one run. The 3.0.0 release run measured 96.4% over ten fixed questions on one model routing, graded by an independent model. It measures whether each claim matches the evidence it cites, not whether the claim is true. (What was measured)
• Google limits the default models. The default routing is Gemini 2.5, and Google now serves each 2.5 model only to accounts that have used it before; on a key from an account that has not, the first call to that model is refused. Choose other models in Settings → Models.
• Custom agent instructions are not private from a bundle's readers. A run's bundle carries them in full. See Agent instructions.
• Desktop builds are unsigned, so macOS and Windows both warn on first launch. The macOS build is for Apple Silicon only, and the demo it prepares on first launch fails on a fresh install — see Desktop app.

Where to go next