{"full_name":"callstack/agent-device","owner":"callstack","name":"agent-device","shard":"1","url":"https://github.com/callstack/agent-device","description":"Open source device layer for AI coding agents: a CLI, MCP server, and typed Node.js API to control mobile, web, TV, and desktop apps on devices and simulators, locally or in the cloud.","homepage":"https://agent-device.dev","topics":["agentic-ai","agents","automation","mobile","testing","adb","android-emulator","e2e-testing","expo","ios-simulator","mcp","react-native","xcuitest","ai-agents","flutter"],"tag_pages":["adb","agentic-ai","agents","ai-agents","android-emulator","automation","e2e-testing","expo","flutter","mcp","mobile","react-native","testing"],"language":"TypeScript","license":"MIT","avatar":"https://avatars.githubusercontent.com/u/42239399?v=4","created_at":"2026-01-30T18:41:54Z","pushed_at":"2026-11-15T11:09:54Z","metrics":{"stars":5405,"forks":406,"watchers":6,"issues":1222,"prs":3177,"releases":156,"issues_open":111,"issues_closed":1111,"prs_open":28,"prs_closed":162,"prs_merged":2987,"discussions":7},"growth":{"day":{"stars":12,"forks":2,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-14","2026-11-15","2026-11-16"],"stars":[5384,5393,5405],"forks":[403,404,406],"watchers":[6,6,6],"issues_open":[101,109,111],"prs_open":[18,33,28]},"boards":{},"same_owner":[],"issues":[{"number":4368,"title":"Repository audit remediation for 10 November 2026","comments":201,"created_at":"2026-11-10T19:59:11Z"},{"number":4045,"title":"Architecture quality: runtime ownership, collocated teardown, trustworthy impact and test evidence (umbrella)","comments":30,"created_at":"2026-10-29T11:27:45Z"},{"number":4441,"title":"Align artifact transfer deadlines and stall budgets","comments":29,"created_at":"2026-11-11T18:51:27Z"},{"number":4060,"title":"test(provider): fix recording timeout reproduced by the related-test gate on unchanged base","comments":13,"created_at":"2026-10-29T12:01:38Z"},{"number":4456,"title":"Pre-dispatch admission: one owner, budget and evidence path for resolved-target input checks","comments":7,"created_at":"2026-11-12T08:14:46Z"},{"number":4426,"title":"Drive Android boot polling with typed observations instead of error text","comments":7,"created_at":"2026-11-11T11:02:50Z"},{"number":4388,"title":"iOS: resolved presses can hit a fading overlay while reporting success","comments":7,"created_at":"2026-11-10T23:53:15Z"},{"number":4107,"title":"iOS physical device over Wi‑Fi: every tap fails with \"lost its transport response\" while snapshots work (USB works)","comments":7,"created_at":"2026-10-30T18:40:12Z"}],"discussions":[{"number":4170,"title":"(iOS/Android) Foldable devices splitview","comments":0,"category":"Ideas","created_at":"2026-11-02T10:11:16Z"},{"number":3228,"title":"Idea: parity runs — replay one flow in two sessions and compare at named checkpoints","comments":0,"category":"Ideas","created_at":"2026-10-05T11:42:55Z"},{"number":2655,"title":"Interested in a PR for iPhone Action Button support?","comments":1,"category":"Ideas","created_at":"2026-09-17T00:11:46Z"},{"number":2024,"title":"Can connect (BrowserStack) ever produce \"did not finish booting in time\", or is that open-only?","comments":1,"category":"Q&A","created_at":"2026-08-25T09:35:37Z"},{"number":812,"title":"Comparisons with https://mobilenext.ai/#products","comments":1,"category":"Announcements","created_at":"2026-06-16T11:45:55Z"},{"number":638,"title":"Porting to Dart and benchmarking a bit","comments":2,"category":"Ideas","created_at":"2026-05-30T19:49:02Z"},{"number":506,"title":"Welcome to agent-device Discussions!","comments":0,"category":"Announcements","created_at":"2026-05-13T06:14:43Z"}],"latest_release":{"name":"v0.21.37 🍰","tag":"v0.21.37","published_at":"2026-11-12T21:16:54Z","body":"## 🍰 Highlights\n\n- **TV apps on LG webOS and Samsung Tizen**: new optional provider plugins `@agent-device/webos` (#4098) and `@agent-device/tizen` (#4168) drive web apps on TVs. They support `open`, `snapshot`, remote keys, pointer, text, scroll, `screenshot` and `logs`.\n- **Limrun iOS runs on the shared XCTest runner** (#4224, #4182). It uses the same snapshot and interaction paths as a local Simulator. Local app installs sync to the instance (#4225), and boot and shutdown map to instance leases (#4275).\n- **Parallel web sessions** (#4191): each web session is its own device, so sessions with different `--session` names run side by side on one host. Web device ids now include the session (`agent-browser-chrome:session:...`).\n- **Pin a workspace to a device** (#4384) with `devicePin` in `agent-device.json`. Commands select within the pin, and conflicting selectors fail with `device_pin_conflict`.\n- **Safer presses on iOS**: `back` is now a verified left-edge swipe (#4460). A synthesized gesture that may already have run is never repeated (#4474). Press and click wait for a moving target within the request's deadline and cancellation (#4461).\n- **Deterministic iOS screenshots**: `screenshot --screen-mask black` masks the simulator outline for stable visual baselines (#4206), and `--capture-backend runner` selects the XCTest runner capture (#4184).\n- **More accurate snapshots**: iOS 27 system toolbars no longer cover their own controls (#4494). Android alert titles and messages stay in `snapshot -i` (#4493). Empty text fields show their placeholder (#4467). Zero-area iOS nodes are marked unmeasured and point at `scroll --until` (#4495).\n- **macOS context menus** open on a native secondary click (#3962).\n- **Hardening from the November audit** across the daemon, remote transport, replay, capture and providers (#4355, #4358, #4360, #4362, #4363, #4366 and more).\n\n## Breaking changes (MCP)\n\n- **Tool results are text by default** (#4466). `agent-device mcp` returns compact text, with no output schemas and no `structuredContent`. To get the previous structured results and output schemas, start the server with `agent-device mcp --structured`. AI SDK tools keep their structured results.\n- **UI targets are top-level fields** (#4459). `press`, `click`, `fill`, `longpress` and `hover` take `ref` (with an optional `label`), `selector`, or `x` + `y`. `get` takes `ref` or `selector`. The nested `target: { kind, ... }` object is rejected with a message that shows the new form. This also applies to `batch` steps, so update saved `batch --steps` files and `client.batch.run` steps. The typed `client.interactions.*` API is unchanged.\n  ```\n  press {\"ref\":\"@e6\"}    click {\"selector\":\"id=\\\"home/products\\\"\"}    press {\"x\":201,\"y\":160}\n  ```\n- **Operator-only inputs left the tool schemas** (#4468). `tenant`, `runId`, `leaseId` and `androidDeviceAllowlist` are refused as tool input in MCP and AI SDK tools. Set them in the server environment (`AGENT_DEVICE_TENANT`, `AGENT_DEVICE_RUN_ID`, `AGENT_DEVICE_LEASE_ID`) or in config. `tools/list` shrank from 268 KB to 216 KB, and every input schema is now 5 KB or less.\n- **Stricter server arguments and session lock** (#4355). `agent-device mcp` accepts only `--structured`, `--help` and `--version`. Configure the server with `AGENT_DEVICE_*` variables or `AGENT_DEVICE_CONFIG`. The session lock now comes only from the serving process (`AGENT_DEVICE_SESSION`, `AGENT_DEVICE_SESSION_LOCK`, `AGENT_DEVICE_PLATFORM`), never from tool input. The stdio server also no longer stalls on blank or malformed lines, keeps answering `ping` and `tools/list` during a call, and cancels the daemon request on `notifications/cancelled`.\n\n## What's Changed\n* feat(provider-webos): add LG webOS TV provider plugin by @alimek in https://github.com/callstack/agent-device/pull/4098\n* fix(ios): report an unreadable system prompt as ALERT_UNRESOLVED, not absence by @thymikee in https://github.com/callstack/agent-device/pull/4149\n* fix(ios): ","is_prerelease":false},"readme":true,"readme_zh":false,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}