Skip to content

AGENTS.md

Latest commit

 

History

History
72 lines (49 loc) · 4.26 KB

AGENTS.md

File metadata and controls

72 lines (49 loc) · 4.26 KB

AGENTS.md — Kobiton Automate

Cross-tool agent instructions for the Kobiton mobile testing platform's MCP plugin. This file is the host-agnostic equivalent of skills/run-automation-suite/SKILL.md — it's read by Gemini CLI (via contextFileName), Codex CLI, GitHub Copilot CLI, ChatGPT Apps SDK, and other agentic CLIs that don't consume Claude Code's skill format.

What this plugin does

Kobiton is a real-device mobile cloud for Android + iOS testing. This MCP plugin gives AI agents 12 tools to:

  • Devices — list, get status, reserve, terminate reservation
  • Apps — list, upload, confirm upload, get details
  • Sessions — list, get, get artifacts, terminate

The MCP server runs at https://api.kobiton.com/mcp. Authentication is OAuth 2.1 (default) or API key (CI/headless).

When the user asks to run tests on Kobiton

Default workflow (matches the run-automation-suite skill for Claude Code users):

  1. Identify the app — ask the user whether to upload a new app build or reuse an existing one. Do NOT auto-upload without confirmation.
  2. Select a device — call listDevices with the right platform filter. Confirm with the user before reserving.
  3. Parse capabilities — read the local Appium test script (Node / Python / .NET / Java), extract the capabilities literal, reconcile against the selected device per the must-match / suggested-default / user-controlled policy in skills/run-automation-suite/references/capabilities.md.
  4. Confirm and execute — present the summary, get user confirmation, run the script in the background, open the live-view URL.
  5. Collect artifacts — after the session terminates, call getSession + getSessionArtifacts for video, logs, screenshots, test reports. Surface session link + pass/fail.

Detailed step-by-step instructions live in skills/run-automation-suite/SKILL.md — read those if you support Claude Code's skill format.

Known limitations

Several behaviors of the current Kobiton MCP server have known gaps that agents should plan around (full details with workarounds in skills/run-automation-suite/SKILL.md § "Known Limitations"):

  • confirmAppUpload async race — returns 200 OK before the parser finishes. Poll getApp(appId) until state is READY or FAILURE_PARSING before downstream calls.
  • reserveDevice ambiguous conflict — device_unavailable lumps 4 failure modes. Don't retry the same device; broaden the filter and pick a different device.
  • W3C /se/log silently breaks legacy driver.getLogs() — Kobiton's Appium endpoint is W3C-strict. Warn the user if their test script uses the legacy log API.
  • deleteSession ~5min cooldown — device enters cleanup after termination; reserveDevice on the same device may return device_unavailable for ~5min.
  • Per-command session data not exposed — no plugin-side way to save a session as a test case; direct the user to the Kobiton portal manually.

userIntent format

Every tool call requires a userIntent argument summarizing what the user is trying to accomplish. The plugin's audit logging consumes this.

For automated / pipeline use, the strict format is:

[partner=<name>] <verb-phrase, 20-80 chars> | contact:<email>

For interactive end-users, a short natural-language summary is fine:

"reserve a Pixel 7 to run the checkout suite"

Cross-host install

Host Install path
Claude Code /plugin install automate@kobiton (uses .mcp.json + skills/ + agents/ + hooks/)
Gemini CLI gemini extensions install kobiton/automate (uses gemini-extension.json + this AGENTS.md)
Cursor Add .cursor/mcp.json config (see README)
Codex CLI ~/.codex/config.toml (see README)
GitHub Copilot CLI See README install section
ChatGPT Apps SDK Add https://api.kobiton.com/mcp in ChatGPT developer mode
Continue / Cline Add to ~/.continue/config.json or equivalent (see README)

Hooks (the hooks/ directory) and the agents/ directory are Claude Code-specific today; other hosts will ignore them.

Reference