Start recording per-process CPU/memory over time (like screen recording, for metrics). Stop with android_performance_record_stop.
| Name | Required | Description | Default |
|---|---|---|---|
| pkg | No | Focus processes containing this package; else top by CPU | |
| serial | Yes | Android device serial (from `device_list`) | |
| intervalMs | No | Sample interval ms (default 1000) |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries the full burden. It describes the start function but doesn't disclose details like how long recording runs, what happens when device disconnects, whether recordings accumulate if not stopped, data volume implications, or what the output/metrics look like. For a stateful start/stop pair with zero annotation coverage, this is a notable gap.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
One well-structured sentence with a parenthetical analogy and a clear pairing reference to the stop counterpart. Efficient and front-loaded, though the analogy placement is slightly awkward mid-sentence.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a stateful paired tool with no annotations and no output schema, the description should explain what happens after starting (where does data go?), how to retrieve results, and implications of the intervalMs parameter. The reference to the stop sibling helps but leaves the recording lifecycle and result retrieval unexplained. The sibling stop tool name gives some closure but more is needed.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, covering pkg, serial, and intervalMs. The description adds context via the analogy and clarifies pkg focuses process filtering ('Focus processes containing this package'). The description doesn't repeat schema details but complements them modestly. Baseline 3 with slight bonus for the pkg clarification.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the verb (Start) and resource (recording per-process CPU/memory over time), with a helpful screen-recording analogy. It distinguishes from android_performance_snapshot by emphasizing time-based recording. However, it doesn't explicitly distinguish from the ios_performance_record_start or android_performance_record_stop, though the name largely encapsulates this.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description indicates this is for time-series metrics (like screen recording for metrics) and explicitly references stopping with android_performance_record_stop. The pkg parameter hints at focusing on specific processes. However, it doesn't specify when to prefer this over android_performance_snapshot or when recording would be inappropriate.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Add one secure layer between your agents and this server.
The set contains near-identical duplicate families: web_* and playwright_* expose ~15 pairs of the same desktop-grid-browser operations (web_get_text/playwright_get_text, web_reload/playwright_reload), and screenshot/log/network/mock capabilities each have 5-8 entry points (device_screenshot vs android_mjpeg_screenshot vs ios_screenshot vs ios_fast_screenshot vs web_screenshot vs webpage_screenshot vs session_screenshot). Many individual descriptions carefully draw boundaries (devtools vs traffic, HID vs session), but an agent cannot reliably distinguish web_* from playwright_*, and ios_screenshot/ios_fast_screenshot/ios_mjpeg_screenshot blur together.
The prefix scheme is broken: Android functionality is split arbitrarily between android_* and device_* (device_screenshot vs android_mjpeg_screenshot), the desktop browser gets two parallel prefixes (web_* and playwright_*), and verbs vary across equivalents (device_navigate_url vs web_navigate vs ios_safari_navigate). session_* uses bare verbs (session_url, session_back), and the same concept gets different names (ios_clipboard_get_hid vs ios_get_pasteboard; device_screen vs ios_orientation).
333 tools is an extreme count by any measure — far beyond the 50+ threshold — and much of the bulk is duplicative (the web_*/playwright_* pairs alone double ~15 slots) or out-of-scope for a device-control server (TestRail, Jira, AzDO, agent memory, secret variables, feedback). Even granting that remote device control + test automation is a broad domain, this surface will devastate agent context budgets and is impossible to navigate coherently.
The core device-control and test-automation domain is remarkably thorough: Android and iOS each have full interaction, app-lifecycle, file, network/proxy, performance, crash, accessibility, recording, and replay coverage, with CRUD lifecycles for flows, suites, app uploads, TestRail cases, and visual-review baselines. Minor gaps exist at the margins — Jira/AzDO lack update/transition/comment operations, and iOS cannot open/close tabs — but the central workflows have no dead ends.
