Local-first agent runtime. Turns context into durable work — with your explicit approval.
Caution
Research and exploration. Dosk is under active development: this repo is a workspace for experimenting with local-first agents, capture, and runtime design—not a finished, stability-guaranteed product. Expect rough edges, shifting APIs, and incomplete features. Use it to learn, prototype, and contribute; don’t rely on it for life-critical or compliance-sensitive workloads without your own hardening.
Dosk is an execution system
The core objects are:
- Task — a unit of intent with a goal, constraints, and dependencies
- Run — an execution of a task: steps, tool calls, decisions, outcomes
- Artifact — a durable output: file, draft, checklist, verified result
- Approval — an explicit human gate before side effects happen
- Policy — rules that govern what can run, what stops, what logs
- Recovery — what happens when a run fails, is interrupted, or the machine restarts
- Prediction — contextual task suggestions based on screen activity, time, and patterns — no prompting required
Every surface (bar, workspace, CLI, command menu, operator panel, APIs, MCP) is a thin client of the same runtime loop. They don't invent separate behavior. They all read from and write to the same task model.
See VISION.md for the full philosophy.
Dosk watches your context — screen, calendar, browser, files, learned patterns — and surfaces tasks without you asking. "You've been reading research articles all afternoon — want a summary drafted?" or "Your Downloads folder has 47 unorganized files." You review, approve, Dosk executes.
Anticipation, not autonomy. You stay in control.
Status: beta — prediction engine in crates/dosk-runtime-core, UI panel scaffolded, Tauri command exposure in progress.
Runs are durable. If a run fails, the machine restarts, or you switch surfaces, Dosk picks up mid-task with full context restored. No starting from zero. No reconstructing intent.
Start work in the desktop bar, inspect it in the workspace, approve it from the CLI, resume it later. Every surface reads and writes to the same task model.
Status: in development — bar, workspace, CLI working; voice surface coming soon.
Approvals, audit trails, policies, and recovery aren't features layered on top — they're load-bearing primitives in the runtime.
| Layer | What it does | Status |
|---|---|---|
crates/dosk-runtime-core |
Execution kernel: causal events, versioned state, typed tool dispatch, DAG graph runner, run state machine, resource scheduler | Stable |
crates/screenpipe-engine |
Local capture and API backend — screen, audio, OCR, accessibility | Stable |
crates/screenpipe-audio |
Audio capture and transcription | Stable |
crates/screenpipe-screen |
Screen capture and OCR | Stable |
crates/dosk-noone |
AI agent runtime with task prediction engine | Stable |
apps/dosk-app |
Desktop app (Tauri + Next.js) | Stable |
Prerequisites: Rust (cargo), Bun for JS/TS, platform deps in CONTRIBUTING.md.
# Build the capture backend
cargo build --release --features metal,apple-intelligence
# Desktop app (macOS)
cd apps/dosk-app
bun install
bun tauri build --features metal,apple-intelligenceTests:
cargo test # Rust
bun test # JS/TSKey paths:
apps/dosk-app/— desktop app (Tauri + Next.js)crates/dosk-runtime-core/— execution kernel (events, state, graph, scheduler)crates/screenpipe-engine/— capture and API backendcrates/dosk-noone/— agent runtime with prediction engine
- Solved autonomous execution — we're honest about current limits
- Invisible always-right background execution — approvals are required
- Mandatory cloud — local-first, zero sign-in for core functionality
- Canvas theater — this is not a visual thinking app
Near-term
- Tauri command for predictions (scaffolding done, wiring in progress)
- Calendar integration for meeting prep predictions
- Pattern learning from approval history
- Approval refinement — richer gates, scoped permissions
Medium-term
- Email automation with approval gates
- Meeting recording and recall
- Artifact versioning (diffs, rollbacks)
- Multi-run workflows spanning days or weeks
- Policies as code
Long-term
- Voice-first runtime surface
- Cross-device resumability
- Distributed runtime — multi-user approval, delegation, handoff
| Doc | Purpose |
|---|---|
| VISION.md | Product vision and principles |
| CONTRIBUTING.md | Build, run, submit changes |
| TESTING.md | Regression checklist for sensitive areas |
| CLAUDE.md | Agent and maintainer notes |
See CONTRIBUTING.md. AI-assisted PRs welcome — run cargo test and bun test, and check TESTING.md before touching window management, audio, or capture.