A Desktop Workspace That Grounds AI in Evidence

What it does

This is an Electron desktop app that pulls in local files, repositories, conversations, and even meeting audio, then organizes them into project-specific Context Rooms. Inside a room, versioned sources, a wiki, layered memory, and documents live together so an AI agent can see a bounded, cited view of your work instead of a global data dump. You write, review agent edits, and delegate tasks from a single offline-capable workspace.

The interesting bit

The architecture treats context as evidence first and generation second. Every source gets a content hash and provenance trail; document edits commit to a transactional core before any side effects fan out to knowledge or memory. It is the rare AI-adjacent project that seems more worried about traceability than scale.

Key highlights

• Room-scoped agents: Agents receive only the tools and documents tied to a specific Room, not an unbounded corpus.
• Evidence ledger: Sources are versioned, addressable, and traceable so you can inspect where a summary came from and correct it.
• Local-first stack: Runs as an Electron app with a Fastify gateway and SQLite WAL; optional services and model credentials are not required to start.
• Reviewable agent edits: Context Docs use a Tiptap-based operation kernel, so changes go through a commit-like flow before they stick.
• Reality Perception: Can capture and transcribe user-triggered microphone or system audio, turning real-world conversations into reviewable records.

Caveats

• macOS is the current development target; other platforms are not yet supported.
• The project is under active development, and the README warns that APIs, storage schemas, and agent contracts may still change.
• Connector coverage for external services like Slack, Notion, or Gmail is described as “still expanding.”

Verdict

Worth a look if you are building agentic workflows and tired of watching LLMs confidently reason over an entire hard drive. Skip it if you need a stable, cross-platform production tool today; the README is explicit that it is not finished baking.