BriefLoop
中文 GitHub ↗ Download
Local · Open source (MIT) · v0.31.0

Turn scattered material into
reports you can edit and check.

A local report workbench for industry research, strategy and IR. Research a question, draft with sources, preserve revisions and export to Word.

Download desktop app Read the guides

Other platforms & installation notes

macOS · Windows Python 3.11+ Multiple runtimes · bring your own key Custom API protocols Server listens on localhost only
BriefLoop v3.1 prerelease report editor, source card and market convention (synthetic sample)
v3.1 prerelease UI Actual report editor and source card · synthetic sample; no model call
Saved example

A real report, generated by BriefLoop

An August 2026 Tencent executive reading scenario, based on five public sources. Preserved report text, generation record and review limits. Not an official Tencent report.

Read the saved report
Actual BriefLoop v3.1 editor showing a synthetic weekly report
v3.1 prerelease UI · synthetic sample, not the Tencent report. The saved Tencent report and its review record are linked above; Word native-reader verification remains pending

SAVED REPORT · CHINESE EXCERPT

腾讯2026年8月经营观察月报

AI投入已进入规模化阶段，现金流纪律成为首要约束。2Q2026资本开支528亿元、同比增加176%，同期资本开支付款593亿元，使自由现金流转负为138亿元；若剔除算力采购预付款项，自由现金流为376亿元。投入与当期回款之间的时间差正在放大，月度资源安排应同时盯住付款节奏与预付款规模。[1] [2]

最近披露的经营基线（节选）
指标（未经审核）2Q20261H2026
收入2,0484,012
资本开支528847
经营活动现金净额5271,541
自由现金流（含预付款口径）−138—
自由现金流（剔除算力采购预付款）376—

金额为人民币亿元（百万元÷100）。这是季度与半年度数据，不代表2026年8月单月表现。

Saved outputExcerpt from a saved report; not an official Tencent publicationRead it in full →
Why

Sources, drafts and review in one place.

Keep a local workspace, inspect evidence for important claims, and evaluate a specific draft. Cloud models and search services still receive relevant content.

A local workspace

Workspaces and records are stored locally. Cloud models and search services receive relevant content when used. Disabling search does not stop model data transfers.

Every claim can be questioned

Important claims are pinned to a line of source text, a PDF page, an Excel cell or an image — never a bare “as far as I know”.

Scoring is separate from drafting

Evaluation and review apply to specific draft versions. A low score does not hide the working draft. Reviewer isolation depends on the runtime and release.

How it works

Complex work, coordinated.
Clear reports, for you.

Bring your materials. Get an editable report with evidence you can follow.

  1. 01

    Set the brief

    A question and its materials.

  2. 02

    Research & write

    Evidence into a first draft.

  3. 03

    Review independently

    Check quality and sources.

  4. 04

    Edit & deliver

    Keep revisions. Export Word.

When useful, turn feedback into experience for the next report.

Explore the architectureClose the architectureSee how the agents work together
BRIEFLOOP

One report. Distinct responsibilities.

The main agent researches and revises. Evaluation and evidence checks run independently.

01

Brief & materials

Audience · goals · requirements

  • Uploaded files
  • Public sources · authorised MCP
02

BriefLoop main agent

Plan → research → write

  • Scout · retrieve evidence
  • Analyst · analyse materials
03

Independent review

Fixed draft and source versions

  • Evaluator · report quality
  • Reviewer · read-only evidence checks
04

Edit & deliver

Working drafts stay editable

  • Web revisions · saved versions
  • Word export · source locations

↶ Findings return to the main agent for targeted revision

Feedback & learning

Optional
  1. Human requests / findingsPreserve the original feedback
  2. Wiki MaintainerOrganise reusable experience
  3. Skill ProposerPropose skill candidates
  4. Compare & reviseRecord adoption and rollback

↶ Wiki / Skills inform the next research task

Wiki is experience, not current factual evidence. Explicit human requirements remain; candidate results still need checking.

  • Local workspace · draft & source versions
  • Runtime · models & permissions
  • Task records · saving & recovery

Workspaces stay local; cloud calls use your chosen services. Read-only review depends on the actual runtime.

Traceable delivery

Every important claim can be questioned.

Delivery is not “the model says it is finished” — it means pinning each claim back to the line of source text, the PDF page or the cell it came from.

  • Claim binding Important claims are pinned to a line of source text, PDF page, Excel cell or image — never a bare “as far as I know”. Once the body text changes, the affected bindings are re-checked.
  • Read-only review The Reviewer checks fixed versions of artifacts and sources. Read-only isolation depends on runtime support; a prompt alone does not establish isolation.
  • Audit bundle Export a zip that carries originals, excerpts or pointers only, chosen per source; it verifies offline against structure and consistency, and holds no credentials or hidden reasoning.
  • Version binding The final Word file is bound to its independent review and evidence version; earlier finals are kept, and a correction or update becomes a new version.
BriefLoop source tracing: a citation in the report resolves back to its original source and evidence (synthetic sample data)
v3.1 prerelease UI Citations back to their source · synthetic sample data, no API keys or real business material

How to inspect evidence and review findings

Capabilities

Enough, and transparent at every step.

From chat to final delivery inside one workbench; the activity of public tools and sub-agents stays visible throughout.

Find sources

Upload original files, search public information and use explicitly authorized connector materials.

Sources and data

Write and revise

Keep text, tables and citations in an editable draft. Save revisions and export to Word, PDF, HTML or Excel.

Editing and export

Check and reuse

Inspect review findings and source scope. Turn feedback into candidate experience and check adoption or rollback.

Review and feedback

BriefLoop report body in detail: editable prose, tables and positioned citations (synthetic sample data)
v3.1 prerelease UI Report body · prose, tables and citations edited in one document
Budget and length

Control research volume and report length.

Search quota is shared by Scout across the whole report; the word target and the cap are yours to edit.

Research limits and costs

Set limits for the current report in the app. Controlled retrieval and runtime-native search have different metering boundaries.

Read the cost guide

Write for the reader

Set the target length and cap for the task. Preserve necessary evidence and state gaps rather than filling the report to a quota.

Before you start

Do I need a model subscription?

Actual generation needs an authenticated runtime and an available model. Model and search providers charge for their services; BriefLoop does not include their credits.

Read the guide
Will my materials leave this computer?

Workspaces and records are local. Cloud-model requests may include relevant sources, drafts and feedback. Search and connectors also send requests to their configured services.

Read the guide
Does a high score mean the report is ready to deliver?

A score is one signal for a specific version. Inspect evidence, scope and unresolved findings. A downloadable working draft is not automatically a formally verified deliverable.

Read the guide
GET STARTED

Install, then create your first report.

Apple silicon Mac and Windows x64 installers are available. Use a compatible Python 3.11+ installation and an authenticated model runtime.

Download desktop appCreate your first report

Initial dependency preparation requires a network connection. Model and search usage is billed by your chosen providers. Workspaces are local, but relevant content is sent to cloud services when used.

Source and CLI installation

Make your next report the traceable kind.

Choose the installer for your computer and follow the guides to create an editable report.

Download desktop app or read the install steps
Engineering

Technical report v3

Read how BriefLoop preserves research, revisions and evidence. This report covers the released 0.21.0, including fact checking, scheduled reports and feedback learning, with an evidence index, architecture diagram and recorded verification scope.

Read the report (Chinese) ↗