Lumi

Native desktop AI agent platform written in Rust. It can run with Claude, Crush, GeminiCLI, HERMES, CURSOR Cargo workspace with a GPUI shell on top, a fleet of lumi-* crates for storage / memory / dispatch / policy / audit / pricing / cron / skills underneath, and per-team Claude Code instances driven over MCP for tool calling.

What it is

The user-facing piece is lumi-gpui — a single binary hosting ten task surfaces in a left-rail layout:

• Chat — lead conversation with streamed assistant text, tool-call bubbles, permission modals, slash commands (/clear, /model, /runtime, /perm, /help), file attachments, @-mention typeahead, and live sub-agent status mirroring.
• Cowork — skill-tile launcher that primes the composer with a domain-specific prompt.
• Code — repo browser + diff preview.
• Creative — ComfyUI supervisor controls (start / stop / restart / open).
• Memory — Search, Observations, Episodes, Knowledge (table + force-directed canvas), Drafts; with run-now buttons for the capture / consolidate / distil passes and side-by-side team compare.
• Analytics — per-team turn / token / USD rollups, daily stacked-bar chart, per-Epic table, and an Audit sub-tab with filter pills + substring search.
• Sprint — five-lane kanban grouped by Epic > Story > Task with collapsible epics and a status-pill click model.
• Plans — surfaced plan files with an inline "Open in Plans" prompt above the chat.
• Agents — roster panel for hiring / firing teammates against the active team.
• Settings — theme, working dir, heartbeat, MCP servers (list + add HTTP), security toggles, governance templates, budgets, memory pipeline knobs, state-git init / commit / remote / push, skills library.

Demo

Short clip (~460 KB): apps/lumi-gpui/docs/media/demo-short.mov

Full walkthrough (~16 MB): apps/lumi-gpui/docs/media/demo-full.mov

Running it

From the workspace root:

cargo run -p lumi-gpui

Release build:

cargo build -p lumi-gpui --release
./target/release/lumi-gpui

State lives under ~/.lumi-lyra/:

• teams.json — team registry (default / cowork / creative are system teams; user teams are deletable from the picker).
• gpui-ui-state.json — last-active tab.
• memory/ — per-team observations, episodes, KG triples, wiki drafts.
• audit/ — JSONL audit log read by Analytics → Audit.
• .git/ (optional) — state versioning, managed from Settings → State sync.

Repo layout

Lumi/
├── apps/
│   └── lumi-gpui/          — desktop binary (GPUI + gpui-component)
├── crates/
│   ├── lumi-audit          — JSONL audit log + reader
│   ├── lumi-auth           — auth & token wiring
│   ├── lumi-bridge         — runtime ↔ UI event bridge
│   ├── lumi-browse         — headless browser tool
│   ├── lumi-core           — engine types
│   ├── lumi-cron           — scheduled tasks
│   ├── lumi-dispatch       — sub-agent dispatch registry
│   ├── lumi-egress         — outbound network policy
│   ├── lumi-mcp            — MCP server scaffolding
│   ├── lumi-memory         — observations / episodes / KG / smart recall
│   ├── lumi-policy         — permission & budget engine
│   ├── lumi-pricing        — provider price tables
│   ├── lumi-providers      — Claude / Anthropic providers
│   ├── lumi-skills         — skill loader (seed + user dirs)
│   ├── lumi-storage        — team / session / sprint JSON storage
│   └── lumi-team           — persona merger
├── resources/seed/         — seed personas, skills, templates
└── docs/                   — design + spec docs

The lead Claude Code instance is launched per-team with an MCP config pointing at in-process servers (memory_server, dispatch_server, persona_server, sprint_server, permission_server, skills_server, browse_server, payload_proxy_server, comfy_server). Memory MCP only mounts when the active team has a working_dir set, and the config is rebuilt at conversation start — switch teams and /clear to pick up newly populated memory tools.

Contributing

See CONTRIBUTING.md.