A local-first AI workspace that turns different perspectives into a reviewable plan.

English · 简体中文
Try it · Architecture · Technical tour · Roadmap

Why ChainMind?

A useful AI discussion needs more than another answer: it needs a clear brief, different viewpoints, a human decision, and a result someone can act on.

ChainMind brings streaming chat, role-based discussions, and model comparison into one desktop-oriented workspace. Configure your own provider, inspect each handoff, and keep conversations on your device.

• Discuss in stages. Move from intake through review, task assignment, execution, and a final report.
• Keep a person in the loop. Choose a proposal, provide feedback, and inspect role-specific output.
• Bring your own models. Configure Claude, OpenAI, or Gemini endpoints, including supported compatible gateways.
• Keep context close. Persist conversations locally; desktop integrations add file indexing, tools, MCP, and SQLite-backed administration.

The project is experimental. Desktop integrations have different maturity levels; see the capability boundaries. The main workspace is Chinese-first; the sample walkthrough is available in English and Chinese.

See the workflow

Actual /demo page. Content is authored teaching material, not a live AI run or a quality benchmark.

Example: a three-person team needs an AI code-review assistant. Choose between a local-first design and a shared-cloud design, compare the review and ownership changes, then export the complete discussion as Markdown.

Brief → Review trade-offs → Assign roles → Define acceptance checks → Deliver
                ↑
         You choose the plan

Try it

Requirements: Node.js 22.12+ (22.x), npm, and Git. better-sqlite3 may require a native compiler toolchain; see setup notes. No environment file or API key is needed for the sample.

git clone https://github.com/djblack1209-coder/ChainMind.git
cd ChainMind
npm ci
npm run dev

Open http://127.0.0.1:3000/demo. Choose a plan, step through the discussion, switch languages, and download a report. The sample does not call providers or write to your chat history.

For live use, open http://127.0.0.1:3000/workspace, configure your own API key and base URL, select an available model, and start a conversation or chain discussion. Provider use may incur charges. No provider credentials are bundled; model availability depends on your provider.

To launch the desktop development app:

npm run electron:dev

Electron starts its own Next.js server on 127.0.0.1:3456. There is no verified signed release to download yet. See setup, storage and troubleshooting.

Architecture

flowchart LR
    UI[Next.js + React workspace] --> State[Zustand stores]
    State --> IDB[(IndexedDB conversations)]
    UI --> Client[Unified LLM client]
    Client -->|Browser| API[Next.js route handlers]
    Client -->|Desktop| IPC[Preload bridge + Electron IPC]
    API --> Core[Shared provider adapter + stream parser]
    IPC --> Core
    Core --> Provider[Your model provider]
    IPC --> Local[Files · MCP · Plugins]
    IPC --> DB[(SQLite administration data)]
    Demo[No-key sample walkthrough] --> Sample[Authored content + Markdown export]

Loading

Engineering question	Where to look
How does a stream work in browser and Electron?	lib/llm-client.ts, lib/llm-core.js, electron/llm-proxy.js
How are discussions modeled and restored?	lib/types.ts, stores/chain-store.ts, lib/chain-workflow.ts
How does the UI handle guided collaboration?	components/ChainPanel.tsx, lib/execution-engine.ts
How are native capabilities exposed?	electron/preload.js, electron/ipc-core-handlers.js
How are failures exercised?	tests/llm-client.test.ts, tests/chain-store.test.ts, tests/api-routes-security.test.ts

For a walkthrough of the trade-offs and a five-minute demo script, see the technical tour.

What works and what is still experimental

Surface	Current scope
No-key sample	Interactive, bilingual, two decision paths, Markdown export; authored content
Chat and comparison	Streaming, local sessions, provider settings, export; real calls need your credentials
Guided discussions	Stage and role UI, plan selection, task assignment, retry/stop paths; model quality is not benchmarked
Desktop integrations	File indexing, native tools, SQLite, MCP stdio and plugin modules; not all UI paths are complete
DAG editor	Engine and node components exist; a complete editor is not mounted in the workspace
Cloud collaboration	Not implemented; the sample cloud plan is an architectural example
Distribution	Platform icons and packaging configuration exist; cross-platform installation, signing and update delivery need release work

Local persistence does not mean cloud inference stays on the device. The browser key-storage fallback is not a hardened vault. Keep the server on loopback; this is not a public multi-tenant backend. Details: SECURITY.md.

Development

npm run check:secrets
npm run lint
npm run typecheck
npm run test:ci
npm run build

npm test and test:ci run the same suite. SQLite uses better-sqlite3 13's Node-API binaries, so testing does not automatically rebuild it for a different runtime. See CONTRIBUTING.md.

Roadmap & contributions

The next priorities are a reliable real-provider discussion loop, a smaller guided workflow controller, a real MCP configuration surface, and reproducible desktop releases. Acceptance criteria are in the roadmap. Logo sources and exports are in the brand guide.

Found an onboarding problem or have a useful discussion example? Open an issue. If this direction is useful to you, a Star helps others discover the project.

License & acknowledgements

A root license has not yet been selected. Do not assume MIT or other permissive reuse terms. Licensing and upstream attribution review are tracked as release work.

Built on React, Next.js, Electron, Zustand, React Flow, and the libraries listed in package.json. Desktop administration modules contain GVA-inspired design notes; see attribution status before redistribution.