Adonis0123/adonis-skills / skills/kimi-computer-use
kimi-computer-use
VerifiedInstall, register, diagnose, or upgrade Kimi Computer Use (kimi-cu) MCP for Cursor CLI, Grok, WorkBuddy, Claude Code, Codex, or Hermes; fix kimi-cu tool-call errors such as 'no target app' or 'mode must be full, image, or ax'; or look at and verify a running macOS app UI through kimi-cu, such as checking whether a window matches a design. Not for general desktop automation, messaging, or purchasing, and not a replacement for a host's native Computer Use.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 7.7 KB |
| evals/check_scope.py | script | 5.3 KB |
| evals/evals.json | config | 9.1 KB |
| evals/trigger-eval.json | config | 2.1 KB |
| references/host-configs.md | markdown | 4.3 KB |
| references/maintenance.md | markdown | 2.9 KB |
evals/evals.json
View on GitHub{
"skill_name": "kimi-computer-use",
"evals": [
{
"id": 1,
"name": "install-verify-routine-check",
"prompt": "Check whether Kimi Computer Use is installed and healthy on this Mac. I only want to know if the service is running and whether Accessibility and Screen Recording are granted; nothing else has changed since yesterday.",
"expected_output": "Runs service-status, xpc-ping, and the bundle version read as the routine check, reads permissions only from the xpc-ping output, skips codesign/spctl because no install or upgrade happened, and never probes kimi-cu with --help or runs upgrade.",
"files": [],
"expectations": [
"Runs `kimi-cu service-status` and `kimi-cu xpc-ping` using the full /Applications/KimiCU.app path",
"Reports Accessibility and Screen Recording from the xpc-ping `permissionStatus` line, not from a host doctor command",
"Does not run codesign or spctl for a routine check, or explains they are install/upgrade-time checks",
"Does not run `kimi-cu upgrade` or append `--help` to any kimi-cu subcommand",
"If the app is missing, points to the official Kimi Code plugin flow rather than a reposted binary"
]
},
{
"id": 2,
"name": "register-cursor-cli-merge-and-prove",
"prompt": "Add kimi-cu to my Cursor CLI. ~/.cursor/mcp.json already has a Figma server in it that must keep working. Tell me how you will prove it actually works end to end.",
"expected_output": "Merges a kimi-cu entry with command /Applications/KimiCU.app/Contents/MacOS/kimi-cu and args [mcp, -s, user] into the existing mcpServers object of the user-level file, then proves connection with `cursor-agent mcp list`, discovery with `cursor-agent mcp list-tools kimi-cu`, and model/runtime with the read-only list_apps + get_app_state(com.apple.finder, ax) smoke, reporting the three layers separately.",
"files": [],
"expectations": [
"Uses the exact command `/Applications/KimiCU.app/Contents/MacOS/kimi-cu mcp -s user` and the server name `kimi-cu`",
"Merges into the existing `mcpServers` object and preserves the Figma entry; does not overwrite the file or symlink another host's config",
"Uses `cursor-agent` (not `cursor-cli`) for `mcp list`, `mcp list-tools kimi-cu`, and the smoke run",
"Separates connection, tool discovery, and completed model tool call as three distinct evidence layers",
"Marks the model layer `UNVERIFIED` if the model only prints or describes a tool call",
"Opens only the Cursor CLI section of references/host-configs.md, not other host sections or maintenance.md"
]
},
{
"id": 3,
"name": "diagnose-no-target-app-fast-path",
"prompt": "kimi-cu 报 `no target app`。我在 Hermes 里调 get_app_state 传的是 app=\"Google Chrome\",连着失败三次之后 Hermes 又说 MCP server 'kimi-cu' is unreachable after 3 consecutive failures。是不是权限掉了?要不要重启服务或者重装?",
"expected_output": "Identifies the shown bad app argument and the Hermes client fuse. Uses com.google.Chrome or a discovered pid and a valid mode, waits for cooldown, then retries read-only get_app_state. Stops only after success; a failed retry continues layered diagnosis, and an unavailable retry is UNVERIFIED. Does not infer overall service or permission health from the original errors or immediately restart or reinstall.",
"files": [],
"expectations": [
"Fixes arguments first and retries read-only get_app_state after cooldown; stops only after success, otherwise continues layered diagnosis or reports UNVERIFIED when no retry is possible",
"States that `app` requires a bundle id such as `com.google.Chrome`, not the display name `Google Chrome`",
"Attributes this client fuse to bad calls using the supplied preceding argument errors, without claiming the fuse message alone proves the cause or permission health",
"Does not restart the OS service, reinstall the app, or mass-kill kimi-cu processes",
"Uses the Hermes tool name form `mcp__kimi_cu__get_app_state` or refers to the Hermes section of host-configs.md if host-specific detail is needed"
]
},
{
"id": 4,
"name": "upgrade-and-distribute-evidence-boundaries",
"prompt": "Case 1: Upgrade Kimi Computer Use and then update the kimi-computer-use skill in adonis-skills for six AI hosts; installed skill files are readable but sessions are untested. Case 2: After the upgrade a host reports 11 tools instead of 10 and an old Codex session says Transport closed. Provide the execution approach and evidence boundaries.",
"expected_output": "Uses `kimi-cu upgrade </dev/null` from references/maintenance.md, rechecks version/signature/xpc-ping, proves the new version in a fresh MCP session, treats Transport closed as a reconnect for that host only, accepts 11 tools as a schema change to read rather than a failure, distinguishes repository source from installed distribution from session loading, and marks untested host/model layers UNVERIFIED.",
"files": [],
"expectations": [
"Runs the official upgrade command with `</dev/null` and does not treat EOF as user consent for login or permission prompts",
"Rechecks bundle version, signature, and xpc-ping after upgrade instead of assuming success",
"Treats `Transport closed` as a reason to reconnect the affected host; does not mass-kill kimi-cu processes or rewrite working registrations",
"Does not reject the new version only because it exposes 11 tools; reads the live tool schema instead",
"Distinguishes repository source, installed distribution copies, and whether an existing session loaded the new skill",
"Does not claim real per-host model acceptance from this offline evaluation; marks it UNVERIFIED"
]
},
{
"id": 5,
"name": "ui-check-design-parity-real-tool-call",
"prompt": "用 kimi 看一下当前桌面上的设置窗口是不是按稿子做的",
"expected_output": "Treats this as an in-scope UI check: calls list_apps only if the app identity is unknown, then get_app_state (ax first, image or full when layout matters), compares the observed window with the design, and reports differences. If no kimi-cu tool call completes, reports the check as UNVERIFIED instead of describing the window from code or the design file.",
"files": [],
"expectations": [
"Does not refuse because kimi-cu is already installed; treats the request as a kimi-cu UI check",
"Emits a real `get_app_state` call (with `list_apps` first only when the target app is unknown)",
"Uses a bundle id or pid for `app` and a `mode` of `ax`, `image`, or `full`",
"Bases the parity verdict on the observed app state, not on source code or the design alone",
"Reports `UNVERIFIED` if the turn ends without a completed kimi-cu tool call",
"Stays read-only and does not click, type, or change settings without confirmation"
]
},
{
"id": 6,
"name": "install-cursor-cli-in-scope",
"prompt": "帮我把 Kimi Computer Use 装到 Cursor CLI 上",
"expected_output": "Checks the installed app with the routine service checks, points to the official Kimi Code plugin flow if it is missing, then registers kimi-cu in Cursor CLI per the Cursor CLI section of references/host-configs.md and proves connection, discovery, and a model tool call separately.",
"files": [],
"expectations": [
"Uses the official Kimi Code plugin flow if KimiCU.app is missing, not a reposted binary",
"Registers `/Applications/KimiCU.app/Contents/MacOS/kimi-cu mcp -s user` under the name `kimi-cu`",
"Uses `cursor-agent` for status and the read-only smoke",
"Separates connection, tool discovery, and completed model tool call"
]
},
{
"id": 7,
"name": "corrected-call-fails-and-fuse-cause-unknown",
"prompt": "Offline decision exercise; do not operate the desktop. A kimi-cu get_app_state call failed with no target app. I fixed it to app=com.google.Chrome and mode=ax, but the read-only retry timed out. Hermes now says unreachable after 3 consecutive failures; the other preceding failures are unavailable. Is the service healthy and may we stop? What if the cooldown prevents any retry?",
"expected_output": "Does not stop or claim service health. Continues layered service-status/xpc-ping and host discovery/runtime checks for the corrected timeout. Treats the fuse cause as unresolved without the preceding failure evidence. If cooldown prevents retry, reports runtime UNVERIFIED rather than success. Describes checks only; performs no desktop action.",
"files": [],
"expectations": [
"A corrected call timeout continues layered diagnosis instead of ending the argument-error fast path",
"Does not claim service or permission health based on the original argument error",
"Does not attribute all fuse failures to bad arguments without preceding-call evidence",
"A retry that cannot run leaves runtime verification UNVERIFIED",
"Keeps the exercise offline and does not execute desktop actions"
]
}
]
}
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |