Persistent memory for coding agents as a single Go binary — SQLite and FTS5 behind MCP, an HTTP API, a CLI and a TUI, with no Node, Python or Docker required.
agent-deviceMCP Server
by Callstack
Released
v0.20.10
MIT
Summary
Give a coding agent a live feedback loop on real mobile, TV and desktop apps — inspect, tap, type and capture evidence on iOS, Android and HarmonyOS over MCP or a CLI.
Features
- Token-efficient accessibility snapshots with stable element refs and change diffs
- Drives iOS, tvOS, Android, Android TV, HarmonyOS, Vega OS, macOS, Linux and basic web targets
- Captures screenshots, video, logs, traces, network data, crash details and React profiles as evidence
- Saves working interactions as .ad replay scripts for CI, or exports strict Maestro YAML
- Worktree-scoped sessions and host-local device claims keep parallel agents off each other's devices
- Connects to BrowserStack, AWS Device Farm and Limrun device clouds with the same commands
- Available as a stdio MCP server, a CLI, or a typed Node.js API over one execution path
Installation
Set up this MCP server in your favorite AI agent — copy a ready-made configuration below.
Description
agent-device closes the loop between a coding agent and the app it is editing. Instead of writing a change and hoping it works, the agent opens the app on a simulator, emulator or physical device, reads the screen, drives the UI, and captures evidence a human can review on the pull request.
Why the accessibility tree, not screenshots
The server hands the model a token-efficient accessibility snapshot rather than an image. Elements come back as stable refs (@e7 [text-field] "First name"), and after a --settle command the tool returns a diff of what changed rather than the whole tree again. That keeps a multi-step interaction inside a sane context budget, and it means the agent acts on roles and labels instead of guessing at pixel coordinates. Screenshots, video and traces are still captured — as evidence, or as a fallback when a screen's accessibility data is poor.
What it reaches
iOS and tvOS through XCTest, Android and Android TV through ADB with a snapshot helper, HarmonyOS through HDC and ArkUI uitest, Amazon Vega OS through the Vega Virtual Device, macOS through a local helper, and Linux through AT-SPI. Basic --platform web support runs Vercel's agent-browser inside the same session and replay model. Coverage is not uniform — the newer HarmonyOS and Vega backends implement a subset of commands, and agent-device capabilities --platform <platform> reports what a given target actually supports.
Beyond a single session
Sessions are scoped to the caller's git worktree and host-local device claims stop parallel agents from stealing each other's simulators, which matters once several agents run at once. Working interactions can be saved as .ad replay scripts for CI, or exported as strict Maestro YAML. The same commands drive hosted devices on BrowserStack, AWS Device Farm and Limrun.
Three ways in
The built-in stdio MCP server (agent-device mcp) exposes the installed commands as structured tools over the same execution path as the CLI, so an agent that speaks MCP and one that just shells out get identical behaviour. A typed Node.js API (createAgentDeviceClient()) covers the third case: driving devices from your own orchestration code or as model tools in an agent you build yourself.
Built and maintained by Callstack, MIT-licensed, and requiring Node.js 22.12 or newer (24+ for web automation). Callstack documents production use at Expensify and Shopify.
Related MCP Servers
The official W&B MCP server: query experiment runs, Weave LLM traces, artifacts and registries in natural language, and write findings back as a W&B report.
Replicate's official MCP server: search thousands of hosted models, read their schemas, and run predictions on image, video, audio and language models from inside an agent.
2 views
Hugging Face's official remote MCP server — search and inspect models, datasets and Spaces, browse the Hub filesystem over hf:// URIs, and turn Gradio Spaces into callable tools.
2 views