Cellar
Hybrid computer-use MCP server that automates browsers and native macOS/Linux apps by fusing accessibility trees, Chrome DevTools Protocol, and vision into one structured context. Exposes four composable tools — cel_see, cel_act, cel_think, cel_perceive — that work with any MCP client and run fully offline via Ollama.
README
cellar
We handle the workflows that break when all you have is screenshots.
CEL (Context Execution Layer) is an open-source computer use runtime that fuses accessibility trees, CDP, and vision into one structured perception layer. Where screenshot-only agents guess at pixels, CEL reads what's actually on screen — and knows when things change.
Status: Active development. Core runtime fully functional on macOS. MCP server with 4 composable tools. Linux support available. Windows planned.
Hybrid Runtime: What It Handles That Screenshots Can't
| Scenario | Screenshot Agents | CEL |
|---|---|---|
| Browser → Desktop handoff | Lose track when focus leaves the browser | Cortex detects context shift via a11y, continues in native app |
| Stale state (dynamic content changes between read and act) | Act on where the button was | Freshness model detects staleness, re-reads before acting |
| Ambiguous targets (8 identical "Delete" buttons) | ~12.5% chance of clicking the right one | a11y tree resolves by label, role, and structural context |
| Unintended side effects (unexpected modal/popup) | Get stuck or blindly click through | Cortex catches the side effect, records it, agent recovers |
| Impossible actions (auth-blocked, disabled) | Loop forever or timeout | Escalation ceiling: structured → semantic → vision → terminal stop |
Run these scenarios yourself: ./scripts/demo.sh — see DEMO.md for the full walkthrough.
<p align="center"> <img src="docs/diagrams/cel-mcp-analogy.png" alt="CEL is MCP for Computer Use" width="700"/> </p>
What Makes CEL Different
- Structure-first perception — reads what's actually on screen through OS-level APIs, not what pixels look like. Vision is the fallback, not the foundation.
- Hybrid runtime with strategy router — per-action routing: structured → semantic → vision → refresh → terminal failure. Escalation ceiling prevents infinite loops.
- Continuous awareness — Cortex tracks what changed, not just what's there now. Freshness model (fresh / soft-stale / hard-stale) prevents acting on stale state.
- Works everywhere — browsers, desktop apps, terminals, legacy software. One runtime, not separate products for browser vs. desktop.
- Model-agnostic — works with any LLM. Sends structured text, not screenshots. A local 7B model works for most workflows.
- 200x cheaper — structured context extraction eliminates expensive vision model inference on every step.
The Problem
Agentic computer use — AI that operates software through the UI — is the defining trend in AI. But it does not work reliably yet.
In browsers, agents have the DOM but still produce unstable results because they depend entirely on LLM interpretation. Outside the browser — on desktop apps, terminals, native software — it's far worse. Agents rely on screenshots alone, feeding pixels to vision models and hoping they correctly identify buttons, fields, and values.
Meanwhile, rich structured information already exists on every computer: accessibility trees, native application APIs, network traffic, input events. No tool combines these signals into a standard format that any agent can consume.
MCP solved this problem for tool access. CEL solves it for computer use.
<p align="center"> <img src="docs/diagrams/cel-before-after.png" alt="Computer Use: Today vs With CEL" width="700"/> </p>
The Solution: CEL
CEL (Context Execution Layer) is both a context extraction and execution layer. It fuses five streams into a single structured JSON output with per-element confidence scoring:
| Stream | What it provides |
|---|---|
| Vision | Screen capture + vision model analysis |
| Accessibility tree | Platform APIs (AT-SPI2, AXUIElement, UIA) |
| Native API bridge | App-specific adapters (Excel COM, SAP Scripting, etc.) |
| Input layer | Mouse/keyboard — injected, intercepted, logged, replayable |
| Network layer | Traffic monitoring for state change detection |
The agent calls getContext() and gets structured JSON with confidence scores — regardless of which source provided the data. Then it executes actions through CEL using the same multi-source approach. Workflows become replayable sequences of structured contexts and actions, not brittle screenshot-to-click chains.
Works on any interface: browser, terminal, Finder, Excel, SAP, Bloomberg — any OS, any application.
Unlike screenshot-only approaches that route every action through expensive LLM inference, CEL uses structured sources (accessibility tree, native APIs) first and escalates to vision models only when needed. Faster, cheaper, more predictable — and capable of running fully offline.
Use CEL with Claude Code (MCP)
CEL ships as an MCP server with 4 tools. Connect it to Claude Code, Cursor, or any MCP client:
# Build everything
pnpm install && pnpm -r build
# Build native module (macOS)
cargo build --release -p cel-napi
cp target/release/libcel_napi.dylib cel/cel-napi/cel-napi.darwin-arm64.node
codesign -fs - cel/cel-napi/cel-napi.darwin-arm64.node
Pick an LLM provider — the fastest path is the interactive setup (writes ~/.cellar/config.toml):
cellar init
Options: paste a Gemini / Anthropic / OpenAI API key, or install Gemma 4 E4B locally via Ollama for fully-private runs. If you'd rather configure via .mcp.json directly (see below), skip init.
Configuration hierarchy
Environment variables override ~/.cellar/config.toml, which overrides compiled defaults.
# ~/.cellar/config.toml
[llm]
provider = "gemini" # openai | anthropic | gemini | ollama | compatible
api_key = "your-key"
model = "gemini-2.0-flash"
[audio] # optional — enables audio transcription in the Cortex
whisper_endpoint = "https://api.openai.com/v1/audio/transcriptions"
whisper_api_key = "sk-..."
whisper_model = "whisper-1"
# whisper_language = "en" # ISO 639-1 hint — improves accuracy
Full variable list: docs/api-reference.md.
Add to .mcp.json in your project root:
{
"mcpServers": {
"cellar": {
"command": "node",
"args": ["/path/to/cellar/mcp-server/dist/index.js"],
"env": {
"CEL_LLM_PROVIDER": "gemini",
"CEL_LLM_API_KEY": "your-api-key",
"CEL_LLM_MODEL": "gemini-2.0-flash"
}
}
}
}
Restart Claude Code and you'll have four tools:
| Tool | What it does | Modes/Actions |
|---|---|---|
| cel_see | Read the screen — structured elements with types, labels, bounds, confidence scores | 14 modes |
| cel_act | Click, type, scroll, drag — by coordinates, element ID, or accessibility API | 11 actions + CDP eval |
| cel_think | Plan, remember, track runs, autonomous execution (run_goal) | 16 modes |
| cel_perceive | Always-on perception engine (Cortex) — continuous screen awareness | 7 modes |
On startup, the Cortex boots automatically (screen model is warm before your first call) and Chrome CDP is auto-detected.
See docs/quickstart.md for the full setup guide and docs/mcp-server.md for the complete tool reference.
Current State
Cellar is in prototype phase on macOS. The bar for exit is defined in docs/PROTOTYPE_EXIT_CRITERIA.md; the curated regression suite that gates it lives in eval/prototype-subset/.
Gated today (macOS local):
- Local execution on macOS via AX + CDP + screen capture + input injection
- MCP server with 4 composable tools:
cel_see/cel_act/cel_think/cel_perceive - Cortex — always-on perception with background event streams
- Autonomous execution (
run_goal) over the prototype scenario suite: browser happy-paths, grounding, ambiguity, recovery, browser-to-desktop handoff - CLI entry points:
cellar init(setup) andcellar run-goal "<goal>" - BYOK providers (OpenAI, Anthropic, Gemini) and local Ollama (Gemma 4 E4B default)
- Per-role LLM routing — Planner / Observer / Vision / Validator
Built but outside the prototype exit bar:
- Audio capture + Whisper transcription fused into the Cortex world model
- Embedded SQLite + FTS5 for memory / semantic search
- First-party adapters — Excel, SAP GUI, Bloomberg, MetaTrader
- Recorder, live-view, and the wider benchmarks/ suite (50+ tasks + hybrid scenarios)
- napi-rs Rust ↔ Node.js bridge
Later phase — explicitly not prototype work (see docs/ROADMAP.md):
- Linux accessibility (AT-SPI2) and Windows UI Automation bridges
- Remote worker / Docker image / managed VMs (
cellar-worker/exists in-tree as a preview; not wired into prototype gates) - Managed cloud, control plane, billing
- Production confidence calibration
- Portable context maps, community workflow registry
Architecture
<p align="center"> <img src="docs/diagrams/cel-architecture.png" alt="CEL Architecture" width="700"/> </p>
cellar/
cel/ ← Cortex + perception layer (Rust, Apache 2.0)
cel-accessibility/ ← accessibility bridge (AXUIElement, AT-SPI2)
cel-context/ ← unified context API + multi-source fusion + references
cel-display/ ← screen capture (xcap)
cel-input/ ← input injection (enigo)
cel-vision/ ← vision model integration (multi-provider)
cel-network/ ← traffic monitoring + idle detection
cel-store/ ← embedded SQLite + FTS5 (memory, knowledge)
cel-llm/ ← LLM provider abstraction
cel-planner/ ← LLM-driven observe-plan-act loop
cel-napi/ ← Node.js native bindings (napi-rs)
agent/ ← strategy router + goal runner (TypeScript)
mcp-server/ ← MCP server (4 tools: see/act/think/perceive)
adapters/ ← app-specific adapters (browser, Excel, SAP)
benchmarks/ ← eval harness (50+ tasks + 5 hybrid scenarios)
live-view/ ← real-time debug surface (screen + runtime decisions)
cli/ ← `cellar` CLI
Getting Started
Quickstart — Claude Code (recommended)
See docs/quickstart.md for the full step-by-step guide. The short version:
# 1. Build
pnpm install && pnpm -r build
cargo build --release -p cel-napi
cp target/release/libcel_napi.dylib cel/cel-napi/cel-napi.darwin-arm64.node
codesign -fs - cel/cel-napi/cel-napi.darwin-arm64.node
# 2. Configure .mcp.json (see quickstart for full config)
# 3. Grant Accessibility permissions in System Settings
# 4. Restart Claude Code — tools are ready
Quickstart — see what the agent sees
No Rust build needed. Just Node.js 20+ and pnpm:
pnpm install && pnpm -r build
npx tsx examples/quickstart.ts https://github.com/login
This launches a browser, extracts DOM elements as structured ContextElements with confidence scores, and shows what the LLM planner would receive.
Prerequisites
- Node.js 20+ and pnpm 9+ (TypeScript packages)
- Rust 1.75+ (CEL core, accessibility bridge, native bindings)
- macOS 13+ with Accessibility permissions
- Chrome (optional, for CDP features)
Build
# Build everything
make build
# Or separately
make build-rust # cargo build --workspace
make build-ts # pnpm install && pnpm build
# Run tests
make test
CLI
cellar init # Interactive first-run setup (pick LLM provider or install Gemma 4)
cellar setup # Configure AX + CDP permissions on this machine
cellar context # Show unified context with confidence scores
cellar context --json # Output raw JSON
cellar context --watch # Live-update context in terminal
cellar capture # Capture screenshot to file
cellar action click 500 300 # Click at coordinates
cellar action type "Hello" # Type text
cellar action key Enter # Press a key
cellar action combo Ctrl C # Key combination
cellar mcp # Start MCP server (stdio)
cellar mcp install # Print Claude Desktop config
cellar run <workflow> # Execute a saved workflow
cellar train # Enter training mode
Benchmarks
Hybrid Runtime Scenarios (CEL advantage)
5 scenarios designed to test where multi-source perception matters. Run them: ./scripts/demo.sh
| Scenario | What breaks screenshot agents | CEL metric |
|---|---|---|
| Browser → Desktop handoff | Lose context across app boundary | sideEffectWarnings |
| Stale state (2s shuffle) | Click where button was | staleRecoveries, refreshRoutes |
| Ambiguous targets (8 similar names) | Can't distinguish identical buttons | semanticRoutes |
| Side-effect detection (unexpected modal) | Stuck or blindly proceed | sideEffectWarnings |
| Terminal failure (auth-blocked) | Loop forever | terminalFailures |
General Web Tasks
We also benchmark on 50+ general web tasks against other tools:
| Tool | Approach |
|---|---|
| Cellar | Multi-source fusion (DOM + a11y + vision + network), confidence scoring, incremental updates |
| Anthropic Computer Use | Screenshot-only, pixel-coordinate actions via API |
| Browser-Use (OSS) | Hybrid screenshot + DOM (Python) |
| Browserbase + Stagehand | Cloud CDP + AI SDK |
| Browser-Use Cloud | Managed browser-use + custom model |
<!-- BENCHMARK_RESULTS_START -->
Measured on Apple M-series (arm64, 12 cores, 18GB RAM), April 2026. Hybrid suite: 5 tasks testing browser-desktop handoff, stale state recovery, ambiguous targets, side-effect detection, terminal failure. All local tools use Gemini 2.5 Flash. Computer Use locked to Claude Sonnet.
Benchmark results (April 2026 — Hybrid Suite, 5 tasks)
| Tool | Avg Time | LLM Calls | Cost/Task | Success |
|---|---|---|---|---|
| CEL | 20.8s | 1.4 | $0.0005 | 100% |
| Browser-Use OSS | 23.4s | 3.0 | $0.001 | 100% |
| Stagehand v3 | 35.6s | 18.2 | $0.005 | 20% |
| Computer Use | 36.2s | 6.2 | $0.155 | 100% |
| Browser-Use Cloud | 46.5s | 5.6 | $0.003 | 100% |
CEL vs the field:
| vs | Speed | Cost | Accuracy |
|---|---|---|---|
| Computer Use (Anthropic) | 1.7x faster | 310x cheaper | Same (100%) |
| Browser-Use Cloud | 2.2x faster | 6x cheaper | Same (100%) |
| Stagehand v3 | 1.7x faster | 10x cheaper | 5x better (100% vs 20%) |
| Browser-Use OSS | 1.1x faster | 2x cheaper | Same (100%) |
Why CEL wins:
- 1 LLM call per task — structured context means most tasks extract data in a single pass. Competitors need 3-18 calls.
- $0.0005/task — Gemini Flash + context distillation. At 1000 tasks: CEL $0.50 vs Computer Use $155.
- Structured context is free — 500+ elements extracted in 100-400ms via Rust-native DOM fusion, no LLM required.
- Full Rust execution loop — perceive, plan, execute, verify all in Rust. No FFI in the hot path.
- For building reliable automation (not one-off tasks), structured context is the foundation <!-- BENCHMARK_RESULTS_END -->
See benchmarks/README.md for full methodology, per-task breakdown, and how to reproduce.
Roadmap
The forward plan lives in docs/ROADMAP.md. Two related references:
- docs/deployment.md — topology: local / remote worker / managed cloud, Docker scope, model backends.
- docs/oss-boundary.md — what's OSS vs commercial, mirror strategy.
Contributing
See CONTRIBUTING.md for how to get started, and DEVELOPMENT.md for build instructions and conventions.
We welcome contributions — especially:
- Accessibility bridges (macOS AXUIElement, Windows UI Automation)
- New application adapters — see docs/building-adapters.md
- MCP tool improvements
- Test coverage for platform-specific code
- Documentation and examples
Platform Support
| Platform | Status |
|---|---|
| macOS | Primary platform. AXUIElement bridge, Cortex, MCP server — all fully functional. |
| Linux | AT-SPI2 accessibility bridge working |
| Windows | Planned (UI Automation bridge designed, not yet implemented) |
License
- Everything OSS-destined (
cel/,agent/,cli/,mcp-server/,cellar-worker/,live-view/,recorder/,registry/,docs/,benchmarks/,examples/,e2e/,tests/,box/): Apache License 2.0. - Community adapters (
adapters/): MIT. - Commercial-only (
app/, futurecontrol-plane/,cloud/,billing/): proprietary — not covered by this license.
See docs/oss-boundary.md for the full license map and what stays private.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。