Council Decision Workspace
Put four distinct perspectives through a public review, then turn your approved decision into an actionable brief.
- Source repo
- loveramarois-byte/council-lab
- Stars
- ★ 111
- Last updated
- 2mo ago
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 85/100 · Good
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A product lead weighing a paid-plan launch on a fixed timeline, including user value, validation cost, and stop conditions.
- Not a fit if
- Teams that require real-model requests and submitted context to stay on the device
- Organizations that require notarized macOS or commercially code-signed Windows builds
- Source review
- 85/100 · Good
What does this agent do, and when should you use it?
Council is a local-first decision workspace for macOS and Windows, with a Linux source-run option. Its desktop app pairs a Next.js interface with a local FastAPI service, and its offline demo works without an API key. When connected to CC Switch, OpenAI, or a compatible service, four seats examine a question through either sequential deliberation or isolated first responses followed by comparison. Council waits for human approval before final synthesis and produces a DecisionBrief with reasons, unresolved questions, actions, and stop and reopen conditions. Records are stored on the current computer by default; a real provider receives the question and selected public context and may charge for use.
The user enters a question and chooses a workflow level—Guided, Roundtable, or Deep Dive. Four seats, named Xili, Jie Wen, Gou Ce, and Guan Lan in Chinese, review it in sequence. In continuous deliberation, later seats read and respond to earlier contributions; independent first answers keep seats separate until they are compared publicly. The user can add facts during the discussion, and Council waits for approval before synthesis. It creates a DecisionBrief that distinguishes user-provided material, model inferences, and unverified claims, while retaining disagreements, open questions, actions, and stop and reopen conditions. Council does not show or store hidden model chain of thought. Records stay on the computer by default, and API keys use the operating system credential store; a real provider receives the question and selected public context.
- A product lead weighing a paid-plan launch on a fixed timeline, including user value, validation cost, and stop conditions.
- An engineering lead reviewing a database migration for failure modes, staged validation, rollback, and launch gates.
- A team comparing proposals that have competing viewpoints and needs to retain counterexamples, minority views, and open questions.
- Someone organizing medical information into a symptom timeline, warning signs, and questions for a clinician without asking the tool to diagnose.
- Someone organizing legal or financial material into known facts, information gaps, and questions for a qualified professional.
How do you install or deploy this agent?
For macOS or Windows, download and extract the matching Release archive, then launch the app or script. The current macOS open-source build is not Apple-notarized, so the first launch may be blocked. The Windows build has no commercial code signing and may trigger SmartScreen. Running from source requires Python 3.12+ and Node.js 22+:
git clone https://github.com/loveramarois-byte/council-lab.git
cd council-lab
./setup.sh
./start.shHow do you use this agent?
Open the desktop app and choose Local Demo to try four-seat deliberation offline without an API key. For real models, go to Settings → Model Providers and connect CC Switch, OpenAI, or a compatible service; the provider receives your question and selected public context and may charge for requests. Enter a question, choose a workflow level such as Roundtable, add facts during the discussion if needed, and approve before final synthesis.
What are this agent's strengths and limitations?
- Four seats can respond to one another in sequence or answer independently before public comparison, with disagreements and minority views retained.
- Human approval is the default before synthesis, and high-risk workflows cannot close automatically.
- DecisionBrief structures reasons, unresolved questions, actions, and stop and reopen conditions.
- An offline local demo works without an API key, and records are stored on the computer by default.
- Desktop packages include Python, Node.js, and required dependencies for macOS and Windows.
- Real-model use requires provider setup; requests send the question and selected public context to that provider and may incur charges.
- Running from source requires Python 3.12+ and Node.js 22+; users unfamiliar with development runtimes may prefer the desktop package.
- The current macOS open-source build is not notarized, and the Windows build lacks commercial code signing.
- The documented release acceptance result covers one model, one route, and fixed cases, so it does not establish performance for other providers or model quality.
- A phone is only a remote interface; provider requests, credentials, and the database remain on the computer, and LAN access is not intended for open public Wi-Fi.
How does this agent compare with similar options?
Compared with ordinary chat, Council runs a public review across multiple seats, preserves disagreements, and waits for user approval before synthesis. It also turns the discussion into a DecisionBrief with actions and boundary conditions.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Council Decision Workspace This agent | 85 · Good | Desktop appFree + model costs | ★ 111 | 2mo ago | Python | OpenAI API |
| Orkas | 43 · Major gaps | Desktop appFreemium | ★ 9k | today | JavaScript | Codex · Claude Code · OpenAI API · Claude API |
| NextClaw | 58 · Major gaps | Desktop appFree + model costs | ★ 263 | 1d ago | TypeScript | Codex · Claude Code |
| MyAgents | 57 · Major gaps | Desktop appFree + model costs | ★ 591 | 3d ago | TypeScript | Codex · Claude Code · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The evidence describes loopback-only backend access, OS credential storage, allowlisted redacted diagnostics, independent review for high-risk changes, and documented migration recovery. Human confirmation, data-flow explanations, and rollback controls are strong. Deductions reflect reliance on operators to protect the host and data directory, ordinary HTTP for mobile pairing, unresolved DNS rebinding risk, unknown publisher identity, and user responsibility for verifying sources; these keep least privilege, dependency security, and attribution below full marks.
The CI file lists backend tests, frontend lint and unit tests, builds, desktop launcher checks, and end-to-end flows. Documentation also gives timeout and upstream error handling, retry guidance, and migration recovery. These files do not establish that this revision passed those checks. Availability still depends on external providers, and the README's single-model acceptance result covers only a fixed route and case set, so consistency and dependency availability receive moderate scores.
The README and installation guide address ordinary users, developers, and Linux source workflows, with product, technical, and professional information-organization scenarios. Boundaries for medical, legal, financial, and production tasks are explicit. Deductions reflect that activation is primarily user initiated, with little evidence of precise automatic triggers; mode and provider flexibility still depend on configuration and external services.
The repository provides organized documentation, platform-specific installation and troubleshooting, FAQs, known signing limitations, uninstall and update instructions, and an explicit Apache-2.0 license. Deductions reflect version and naming details spread across several entry points, no verified maintainer identity or long-term maintenance commitment in the supplied evidence, and a security policy that supports only the latest default-branch version.
DecisionBrief separates reasons, open issues, actions, stop conditions, and reopen triggers; it preserves the discussion and lets users add facts, supporting usable decision outputs. Deductions reflect the time and potential provider charges of multiple model calls, no ordinary-use cost comparison, and that the claimed benefit over a single model remains primarily a product claim.
The README, installation and security documents, test structure, CI configuration, and license provide traceable support for concrete claims. The documentation distinguishes public discussion from hidden chain-of-thought and separates user material, model inference, and unverified claims. Deductions reflect incomplete cross-file corroboration for some product claims and an acceptance result limited to one route, one model, and a fixed case set.
- Real providers receive the question and selected context and may charge for requests; confirm they are authorized to receive the material before sending it.
- High-risk and professional-domain outputs are decision support only. The system does not verify source content or professional credentials online and does not replace qualified professionals.
- Mobile pairing is not intended for public Wi-Fi. Release packages are not notarized or commercially signed; verify their source and checksums before installation.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.18.9 Council v0.18.9
Release notes excerpt
Council v0.18.9 更新内容 Improved - 所有新生成的结构化决策简报现在都稳定包含带成功标准的下一步行动、明确停止条件和重新审议条件；即使结论可以推进，也先要求可逆、可观察的首步。 - 高风险识别补充常见中文用药、劳动争议和投资表达，减少专业问题因口语化描述绕过高风险控制的可能。 - 完成页与 Markdown、HTML 导出同步展示停止条件，让软件内结果与 GitHub 功能说明保持一致。 Fixed - 修正 App Store 更新端到端测试的错误断言，确保商店分发版本不会出现 GitHub 直发版的“下载并安装”入口。 Testing - 通过 382 项后端测试、61 项 Chromium 端到端测试、12 项前端单元测试、TypeScript 检查、Next.js 生产构建和隔离数据实机流程演练。 安装与升级 - macOS：下载 macOS ZIP，解压后双击 Council.app。 - Windows：下载 Windows ZIP，完整解压后双击 Start Council.cmd。 - 手机端：电脑启动后打开“设置 → 手机连接”，在同一可信 Wi-Fi 下扫码配对。 - 两个安装包均内置运行环境，不需要另行安装 Python 或 Node.js。 …View on GitHub ↗ -
v0.18.8 Council v0.18.8
Release notes excerpt
Council v0.18.8 更新内容 Improved - 审议席位增加连续推进路径，当前席以克制的呼吸动效强调，五席到最终答案的进程更直观。 - 公开讨论增加贯穿发言的审议脊线，并统一等待状态的排版节奏，让长篇讨论更像一份连续卷宗而非消息列表。 - 手机端同步适配新的席位路径与发言脊线，保持无横向滚动，并支持“减少动态效果”系统设置。 安装与升级 - macOS：下载 macOS ZIP，解压后双击 Council.app。 - Windows：下载 Windows ZIP，完整解压后双击 Start Council.cmd。 - 手机端：电脑启动后打开“设置 → 手机连接”，在同一可信 Wi-Fi 下扫码配对。 - 两个安装包均内置运行环境，不需要另行安装 Python 或 Node.js。 - 当前开源构建未做 Apple notarization 或 Windows 商业代码签名，请仅从本仓库 Release 下载。 Full Changelog: https://github.com/loveramarois-byte/council-lab/compare/v0.18.7...v0.18.8View on GitHub ↗ -
v0.18.7 Council v0.18.7
Release notes excerpt
Council v0.18.7 更新内容 Improved - 首页审议谱系补充“分析、追问、构建、反观、结论”短标签，保留移动端紧凑符号布局，让第一次使用的人也能理解五步流程。 安装与升级 - macOS：下载 macOS ZIP，解压后双击 Council.app。 - Windows：下载 Windows ZIP，完整解压后双击 Start Council.cmd。 - 手机端：电脑启动后打开“设置 → 手机连接”，在同一可信 Wi-Fi 下扫码配对。 - 两个安装包均内置运行环境，不需要另行安装 Python 或 Node.js。 - 当前开源构建未做 Apple notarization 或 Windows 商业代码签名，请仅从本仓库 Release 下载。 Full Changelog: https://github.com/loveramarois-byte/council-lab/compare/v0.18.6...v0.18.7View on GitHub ↗