BriefLoop turns a local workspace into a delivery pipeline for reports: you give it the goal and the material, it researches, writes and evaluates independently, and your edits settle into experience for the next run.
Every step leaves a record you can inspect, and no step reports “queued” as “done”.
Title, goal, reader, time range and word target live in one form; upload material, or simply switch search on and let the host go find public sources. Sections meant to be filled in by hand stay marked “to be filled”.
The research budget is shared across this report; Scout decides what to search and which originals to keep, and Analyst writes inside those sources, with citations back to a specific source and location.
The evaluation does not read the author’s self-assessment, and the score binds to the version it scored. Where it helps, one automatic revision and re-check can be scheduled; whatever is left goes back to you.
Word is generated from a frozen version and reports its progress; final delivery binds that version, the template, the sources and the review record. Your edits and feedback are curated by the Wiki Maintainer, and a skill is enabled only once it has been validated.
Traceability comes from the originals, locations and version relations the workspace keeps — not from any single model answer.
Scores, Word exports and final documents all bind to a specific draft version; further edits create a new version, while earlier versions and already-released documents stay readable.
Originals, extracted text, locations and dates live in the workspace; model summaries, search snippets and scores are not evidence.
“Data and gaps” shows metrics, calculations and gaps from what is stored, without calling a model; whatever is missing is listed as missing.
You choose the export scope; the bundle keeps material locations and a record of how each issue was handled, and credentials and hidden reasoning are not part of the export.
The one who writes does not score itself; scoring, curating experience and proposing skills each stand apart.
Talks with you, plans the research and schedules drafting and revision, running on the main-chain model. It is the party you talk to inside the workbench.
Searches, fetches and keeps public sources within the shared budget.
Writes inside the sources, with citations back to a specific location.
Scores in a separate session and compares whether a candidate is genuinely better.
Turns edits and feedback into the workspace Wiki.
Proposes or improves skills from the Wiki, enabled only after they are validated.
The budget belongs to one report and is shared by every scouting run inside it; permissions belong to the host CLI, and BriefLoop does not pretend to exceed them.
12 search attempts · 60 candidate URLs · 18 full-text sources
All three numbers are editable, and 0 means no extra allowance of that kind. A failed search still counts; once the cap is reached the material already gathered is kept and the gaps are listed, and the allowance is not topped up automatically.
30 search attempts · 200 candidate URLs · 45 full-text sources
Recurring industry reports share the same budget; a direct fetch, a fallback extraction and a retry of the same URL together count as one full-text source.
Can be forced into a read-only or read-write sandbox, with online search controlled per turn.
Read-only through permission rules; the model is written as provider/model.
Claude Code, Kimi, Hermes, Reasonix and MiMo run under their own permission models; turning on “online” releases their own web tools.
Python owns storage, tools and deterministic computation; the host CLI owns model execution, and the two meet through the existing jobs/events/Store.
git clone then ./start.sh creates the virtual environment, installs dependencies and opens the web app; there is no second repository to clone.
Implement the bridge protocol, add one catalog entry and declare the capabilities (permission_modes, steer, images); the input area then follows those capabilities automatically.
npm ci && npm run build bundles the frontend and the bridge, and pytest runs behaviour tests; the tests call neither real models nor paid search.
Clone the repository, connect one authenticated host CLI, and start.
