vesper-ios — Claude Code skill: what it does, how to install
vesper-ios is a Claude Code skill that automates iOS reverse engineering, tweak deployments, and runtime debugging. It coordinates Frida, LLDB, and UI automation over SSH.
MIT ★ 83 updated 2026-10-25
npx skills add Wtrwx/vesper --skill vesper-ios -a claude-codeWhat vesper-ios does
This skill drives iOS binary inspection and app automation through a connected vesper MCP server. It batches UI actions for UIKit and SwiftUI views, queries runtime values without heap scans, and captures accessibility hierarchy snapshots.
For reverse engineering, the workflow links binary decryption with classdumpc header extraction or IDA decompilation. Debugging relies on Frida for stack inspections and registers, switching to LLDB frames when needed.
Tweak development follows a minimal cycle: compile, deploy over SSH, restart the application with uiopen, and verify UI changes. Operation requires a running vesper MCP server and an SSH connection to the iOS device.
What it helps with
- automate interactions in UIKit and SwiftUI views
- dump Objective-C headers from decrypted binaries
- inspect memory, registers, and frames via Frida or LLDB
- build, install, and restart iOS tweaks over SSH
Example prompts for the agent
- Run ui_capture on com.example.app and tap the login button
- Dump Objective-C headers for the decrypted binary using classdumpc
- Deploy the tweak over SSH, restart the app, and check logs
How to install vesper-ios
1Skills CLI (Vercel)
npx skills add Wtrwx/vesper --skill vesper-ios -a claude-code2Manually: put the files in place
~/.claude/skills/vesper-ios/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/vesper-ios/
1Skills CLI (Vercel)
npx skills add Wtrwx/vesper --skill vesper-ios -a codex2Manually: put the files in place
.agents/skills/vesper-ios/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add Wtrwx/vesper --skill vesper-ios -a cursor2Manually: put the files in place
.cursor/skills/vesper-ios/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install Wtrwx/vesper skills/vesper-ios2Manually: put the files in place
.github/skills/vesper-ios/SKILL.md1Skills CLI (Vercel)
npx skills add Wtrwx/vesper --skill vesper-ios -a gemini-cli2Platform CLI
gemini skills install https://github.com/Wtrwx/vesper --path skills/vesper-iosManually: put the files in place
.devin/skills/vesper-ios/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about vesper-ios
How do I install vesper-ios in Claude Code?
Run npx skills add Wtrwx/vesper --skill vesper-ios -a claude-code in your terminal. This registers the skill definition in Claude Code.
What does vesper-ios do?
It interfaces with the vesper MCP server to automate iOS app interactions, inspect runtime states, and manage tweak installations.
What is required to run vesper-ios?
You need a running vesper MCP server, SSH access to the target iOS device, Frida, and optionally LLDB, classdumpc, or IDA for analysis.
Source text
SKILL.md · MIT
---
name: vesper-ios
description: Use Vesper for iOS reverse engineering, UIKit/SwiftUI automation, Swift and native runtime values, decryption plus classdumpc/IDA, SSH tweak builds and installation, and LLDB/Frida/log debugging.
---
# Vesper iOS
Use the connected `vesper` MCP server and reuse the configured app session.
Load advanced schemas only when needed with `platform_tool(name="…", action="schema")`,
then call `platform_tool(name="…", arguments={…})`.
## Default workflow
只做最小必要检查。不要创建备份,不要增加或运行额外的本地测试、冒烟测试、长回归或测试报告。
编译成功后直接安装,通过 SSH 重启目标 App,再用 Frida 操作 App,实测当前修改即可。
安装命令成功即视为安装完成;不要再核对进程中加载的模块、构建标记、包版本或逐字节哈希。
能由 AI 完成的编译、安装、重启、点击、返回、输入、滑动和读日志,都由 AI 完成,不要转交用户手动操作。
复用已读信息,只取相关差异和必要行。默认只查目标控件,截图限于复现、修复后和异常节点。
首次关键失败立即停止并修正。只有安装异常、明确怀疑旧版本或出现崩溃症状时,才做对应排障。
用户验收后直接完成已授权的提交、推送,不追加检查,不持续跟踪普通推送的 CI;查询时使用完整 SHA。
没有新改动、新失败或用户明确要求,不扩展测试和收尾流程。
For tweak work: build → install directly → `app_restart` → minimal Frida/UI
exercise of the requested change. Do not create backups or add local test suites,
smoke tests, broad regressions, loaded-module comparisons, build-marker checks
or post-install hash comparisons. A successful install command completes installation.
Do not ask the user to operate the app when the available tools can do it.
`app_restart(device_id, bundle_id)` restarts through SSH `uiopen` and returns a
fresh Frida session; it falls back to the existing Frida launcher if `uiopen` is
missing. Deployment with `restart=true` includes this step. For local or SSH
build commands, read [references/ssh-tweaks.md](references/ssh-tweaks.md).
## Observe and operate with few calls
- `ui_capture(device_id, bundle_id, mode="summary")` finds/reuses the session.
Text requires `include_text=true`; images require `mode="screenshot"` or `"both"`.
- Read subsequent `ui_snapshot` pages with the returned cursor. Omit
`snapshot_id` for a fresh observation after changing the UI.
- Batch related `tap`, `back`, `input_text` and `scroll` steps in `ui_automate`.
Prefer observed view IDs or unique selectors and one final `after_capture`.
An uncertain timeout may follow a completed action: observe before retrying.
- For a paged feed, `scroll` with `pages=1` advances and `pages=-1` goes back.
Reuse the scroll view ID; add `animated`/`settle_ms` only when necessary.
- `objc_query_instances(object_ids=[…], field_names=[…])` reads selected values
without scanning the heap. Use class-name discovery only when no handle exists.
Request controllers, constraints, targets or nested values only as needed.
UIKit operations and object cleanup run on the main thread. `backend="auto"`
supplements SwiftUI hosting views with accessibility nodes; `"accessibility"`
selects the semantic tree and its activate/adjust/escape/scroll actions.
For pure Swift values/calls/hooks use advanced `swift_runtime`; for C/C++ use
`runtime_native` with explicit types/offsets. Both reuse the app session and return
bounded selected data. Swift is lazy and scoped to chosen Mach-O images.
`swipe`, `long_press` and multifinger `gesture` use optional SSH ZXTouch input;
raw coordinates need no Frida attach. Missing input support returns installation
instructions through `input_setup`. Read
[references/swift-native-input.md](references/swift-native-input.md) only when
these backends are needed. Keep the existing Frida deployment; do not add VNC.
## Metadata and analysis
`objc_dump(binary_path=…)` uses host **classdumpc** for offline headers;
`objc_dump(device_id=…, bundle_id=…, class_name=…)` uses runtime metadata.
Use runtime queries for current instance values. Keep headers local and request
only relevant classes or fields.
For encrypted images, chain `app_decrypt` → classdumpc for headers or
`app_decrypt` → IDA/Hex-Rays for implementation logic. Read
[references/analysis-pipelines.md](references/analysis-pipelines.md) when needed.
## Debug and logs
Use `debug_inspect` with the default Frida backend for stacks/registers, or
`backend="lldb"` for stopped frames/locals. If a required component is missing,
use `debug_setup(action="check")` and its installation hints; matching local
Apple support can supply debugserver through `action="install"`.
Start `app_logs` once, filter by cursor/level/text, and stop it when finished.
If LLDB cannot attach, continue with Frida and logs instead of repeated setup.
Read [references/debugging.md](references/debugging.md) only for a backend failure.
Return concise findings and native MCP images; keep full dumps/logs local.
© Wtrwx · MIT · Wtrwx/vesper