agent-device — Watchlist

MCP100 Assessment

Watchlist / Tracking

Why we’re tracking it

Follow-up value is specific to Agent-controlled mobile and desktop app automation with a typed MCP interface: Do verification results and device scopes remain consistent across mobile, desktop and web targets?

Reviewed Oct 5, 2026 · Documentation review; complete setup workflow not independently tested

Best fit

AI-agent verification and automation across mobile, desktop and web apps

What we know

Callstack's CLI, MCP server and Node API for inspecting, controlling and verifying apps across iOS, Android, web, macOS, Linux and other targets.

Maintenance / activity evidence

agent-device 0.21.20 published 2026-10-02T17:51:17.168Z

Tracking evidence observed Oct 5, 2026. These dated signals do not establish current maintenance. Reviewed enrichment confidence: high.

Verification gaps / watch-outs

Installation and comprehensive security have not been independently audited.
Structured device/app automation commands; no generic shell through MCP. Writes: UI interaction; App launch/close; Input actions.
Authentication not yet verified.
Sensitive access: App sessions, Device UI, Screenshots

Structured commands reduce arbitrary-shell risk, but the server can operate authenticated apps and real devices.

Full authority, access and evidence limits are available in Review context and Access & risk.

Popularity

· 554.7k package downloads / 30 days

Downloads observed Sep 25, 2026.

Preserve any already-reviewed adoption metrics in MCP100. Batch 4 did not refresh package downloads because the package-stats provider had insufficient credits.

All limitations / Review context

Do verification results and device scopes remain consistent across mobile, desktop and web targets?

The bounded review of the publisher's documented Agent-controlled mobile and desktop app automation with a typed MCP interface and release notices identified no disclosed material condition preventing an accurately scoped public tracking entry. The resource's operational permissions remain uncertainty; this assessment does not certify installation safety.

Structured commands reduce arbitrary-shell risk, but the server can operate authenticated apps and real devices.
Structured device/app automation commands; no generic shell through MCP.
App UI, accessibility trees, screenshots and device state are exposed to the agent.
Reviewed write actions: UI interaction; App launch/close; Input actions
Sensitive access: App sessions, Device UI, Screenshots
Installation, client compatibility and comprehensive security have not been independently audited. Authentication and access depend on the configured services.
Runtime behavior, dependencies and undisclosed vulnerabilities were not independently audited.
Documented tool authority and any remote data handling must be checked before installation or use.
Tool inventory, client matrix and permissions completeness were not requirements for this Radar tracking review.
MCP intentionally exposes structured automation commands, not a generic shell.
Platform/device permissions and remote clouds vary.

Original activity observations

agent-device 0.21.20 published 2026-10-02T17:51:17.168Z

Technical identity

npm/agent-device

Exact implementation source: https://www.npmjs.com/package/agent-device ↗
Reviewed resource type: Testing / Debugging
Repository identity: callstack/agent-device
Package identity: agent-device

Evidence & Sources

Evidence basis: Recorded source evidence
Setup runtime verification: Not independently runtime-tested by MCP100 for this setup workflow.

Reviewed Oct 5, 2026 · Latest source observation Oct 6, 2026

Package registry metadata ↗
agent-device
Recorded package identity, distribution or download evidence.
GitHub repository ↗
callstack/agent-device
Repository reference; implementation attribution is not established for every document.
Includes references whose attribution is not established.

Setup · Watchlist

Testing / Debugging setup

Publisher instructions.

Source evidence observed Oct 14, 2026.

Additional setup details required — see the specific gaps below.

Warning

Not runtime tested, no third-party installation, network authorization, paid calls, or mutation executed.
Do not claim production service is available based on README alone.
Provides MCP config in README; CLI installation alone does not register client; device automation has sensitive effects
V1 omitted actual stdio MCP subcommand `mcp`.
Official npm publisher package page README inspected for this exact npm identity; static research only. Some setup statements are conservatively summarized; no source code or live client execution checked. V2 review did not run any live client.
Avoid write, payment, token transfer, publishing, sending, booking and remote account changes as tests.
Replace credential, action and registration placeholders; follow the matching publisher/client prerequisites. No execution performed.
Use the matching host, actual registration name and original installation environment; this does not revoke credentials.

Set up testing / debugging integration
Publisher instructions
Provides MCP config in README; CLI installation alone does not register client; device automation has sensitive effects | Publisher npm README revisited on 2026-10-14; exact identity retained.
Reviewed setup example · Publisher instructions
npm install -g agent-device@latest
Reviewed setup example · Publisher instructions
agent-device mcp
Configure / authenticate
MCP100 setup guide
Run `agent-device doctor`. To expose MCP, register command `agent-device mcp` in the chosen client; CLI installation alone does not register MCP.
Previously reviewed configuration / prerequisites: After installing agent-device, configure MCP client with command "agent-device", args ["mcp"]. Device simulator/emulator dependencies and permission prompts may apply.
Check connection
Suggested check
Run `agent-device doctor` and inspect client tools; avoid controlling any device just for validation.
Remove / uninstall / revoke
MCP100 setup guide
If your Claude Code MCP registration is named `agent-device`, run the client removal command. For Claude Desktop/Cursor/Codex or another registered name, remove the matching entry in that host’s MCP settings. Hosted OAuth disconnect and stopping a self-hosted process are separate; npx has no permanent global package install.
Remove named client registration · Claude Code only · MCP100 composed guide
claude mcp remove agent-device

Official Documentation & Sources

What it does

Callstack's CLI, MCP server and Node API for inspecting, controlling and verifying apps across iOS, Android, web, macOS, Linux and other targets.

Capability: Developer Tools
Resource type: Testing / Debugging
Industry: Software & Technology
Best fit: AI-agent verification and automation across mobile, desktop and web apps

Representative uses

Open and inspect apps
Use accessibility snapshots
Press/type/scroll and verify state
Capture screenshots/video/evidence

Limitations

MCP intentionally exposes structured automation commands, not a generic shell.
Platform/device permissions and remote clouds vary.

Reviewed access

Access & risk

Review the exact access boundary before connecting this MCP to accounts, credentials, or sensitive data.

Authentication: Not documented in the reviewed sources
Permissions: Structured device/app automation commands; no generic shell through MCP.
Data handling: App UI, accessibility trees, screenshots and device state are exposed to the agent.
Cost: free
Risk and cautions: Structured commands reduce arbitrary-shell risk, but the server can operate authenticated apps and real devices.

Access scope and cautions

Write actions: UI interaction App launch/close Input actions
Sensitive access: App sessions Device UI Screenshots
Risk level: high

FAQ

Based on reviewed evidence as of 2026-10-06. Watchlist tracking is not a recommendation or a security certification.

Which tasks are documented for this implementation?

The reviewed tasks include: Open and inspect apps; Use accessibility snapshots; Press/type/scroll and verify state; Capture screenshots/video/evidence.

What setup information is documented?

Recorded installation/entrypoint step: agent-device mcp. Recorded transport: stdio. Prerequisites: agent-device installed; Platform/device tooling as needed. These are the documented setup details, not a claim that installation or client compatibility was independently tested.

Can it modify data or perform write operations?

Documented write actions include: UI interaction; App launch/close; Input actions. Recorded permission boundary: Structured device/app automation commands; no generic shell through MCP.

What data-handling boundary should I consider?

App UI, accessibility trees, screenshots and device state are exposed to the agent. Recorded sensitive-access context: App sessions; Device UI; Screenshots.

What limitations and risks should I check before use?

Structured commands reduce arbitrary-shell risk, but the server can operate authenticated apps and real devices. MCP intentionally exposes structured automation commands, not a generic shell. Platform/device permissions and remote clouds vary. Runtime behavior, dependencies and undisclosed vulnerabilities were not independently audited. Documented tool authority and any remote data handling must be checked before installation or use.

Is this an official MCP?

Callstack is listed as First-party MCP in MCP100’s current evidence.