Inspect, automate, analyze and build — from your agent.

English · 简体中文

Quick start · Client setup · Tool reference · Demos · Issues

Vesper is an iOS reverse engineering platform combining structured MCP tools and agent skills for runtime inspection, app automation, decryption, debugging and tweak development. Connect your MCP client to SSH, Frida and an optional licensed IDA installation, then work with the target app through your agent.

The vesper-ios Agent Skill composes these primitives into inspection, decryption + classdumpc/IDA, SSH tweak development and debugging workflows. Codex, Claude Code and Cursor setup examples are included.

See the tool reference for the complete catalog and required fields.

Agent skills compose workflows over MCP primitives. SSH and Frida connect those tools to the target app; classdumpc, IDA/Hex-Rays, LLDB/debugserver, Theos and ZXTouch provide optional analysis, development and physical-input integrations. Swift support is experimental. See automation and analysis pipelines for backend requirements and limits.

The Python package includes the compiled Frida agent. Basic SSH, UI and runtime inspection does not require Node.js, Theos or IDA. Rootless jailbreaks are the primary target. ARM64e tweak builds require an appropriate Apple toolchain on macOS. SDKs, proprietary tools and device services are supplied by the operator.

From the repository root:

uv sync --frozen --all-extras --no-editable
cp -n examples/vesper.toml vesper.local.toml
chmod 600 vesper.local.toml

The lockfile provides reproducible Python dependencies. A non-editable install also works in managed environments that do not load editable .pth files.

Edit vesper.local.toml. Set an absolute workspace path, the SSH endpoint, your key, verified SSH host keys, allowed app bundle IDs and the Frida transport.

workspace = "/absolute/path/to/vesper"
state_dir = ".vesper"

[devices.test]
host = "127.0.0.1"
port = 2222
username = "root"
client_keys = ["~/.ssh/id_ed25519"]
known_hosts = "~/.ssh/known_hosts"
allowed_apps = ["dev.iosdecompiler.fixture"]
frida_device_id = "replace-with-observed-device-id"

The example endpoint assumes an existing local SSH forward; use your device's actual endpoint. Verify and pin the SSH host key before connecting. Select the exact device ID returned by frida_list_devices. For an existing secured Frida SSH tunnel, replace frida_device_id with frida_remote = "127.0.0.1:27043"; configure one transport per device.

Select a config with --config or the VESPER_CONFIG environment variable. --config takes precedence. State defaults to .vesper under the workspace.

For tweak builds, set THEOS or the config's theos path. For IDA, configure ida_directory and ida_python after completing the vendor's normal licensing and Python setup. See the comments in the example config.

.venv/bin/vesper --config /absolute/path/to/vesper/vesper.local.toml

Configure your MCP client to launch this command with absolute paths. Stdout carries the MCP protocol; diagnostics go to stderr. See client setup for Codex, Claude Code and Cursor examples.

The default tool_profile="focused" advertises 20 common tools. Advanced primitives remain available through platform_tool: request one schema with action="schema", then call it with arguments. tool_profile="full" advertises all 76 primitives. Copy skills/vesper-ios into your agent skill directory (for Codex, ~/.codex/skills/vesper-ios) to load workflows on demand.

Structured results are returned once in structuredContent, with a short text status. For clients that only consume text, set legacy_text_json = true to restore the duplicate JSON text. Images always use native image content.

Idle sessions are reclaimed after 300 seconds by default; pending operations, hooks and running workflows protect them. Set session_idle_timeout = 0 to disable this. ssh_probe_ttl, snapshot_ttl, snapshot_max_count, snapshot_max_bytes, object_handle_ttl and maintenance_interval are configurable in the top-level TOML settings. device_info(force_refresh=true) bypasses the SSH probe cache; connection failures/reconnections invalidate it automatically. screenshot_transport = "binary" avoids base64 on the Frida link while preserving the original PNG/scale. Use "base64" for transport compatibility. MCP image content remains native in either mode.

Run one server per state directory. A config without devices supports host-only tools such as server_status. The server itself does not require a model API key.

Optional header dumping uses host classdumpc (PATH or the classdumpc setting). Optional stopped inspection needs host LLDB and device debugserver; debug_setup checks components and installs missing debugserver from local Apple iOS support. See decryption and analysis pipelines for runtime and offline workflows.

Open an allowlisted app on the unlocked phone, then call ui_capture:

{"device_id":"test","bundle_id":"dev.iosdecompiler.fixture","mode":"summary","include_text":false}

mode is required: summary returns visible controls (including custom touch or gesture views), screenshot returns a native MCP PNG image and local_path, and both returns both. Text still requires include_text: true. A valid session is reused; cold capture probes SSH and verifies the running app before attaching. Consecutive captures are live; they do not enumerate apps or attach again while the same verified process session remains valid. No app launch/restart is implicit.

Summary defaults to IDs, classes, frames, interaction flags and parent/depth links. format: "detail", fields, visible_only: false and controls_only: false opt into more data. Controllers, targets and constraints are fetched only when requested. Visibility includes hidden/transparent ancestors and window intersection; occlusion by other views is not inferred.

The summary contains session_id at the capture root and snapshot_id, captured_at, truncated and next_cursor inside summary. Pass the session ID and the cursor to ui_snapshot to read the next immutable page without accessing the device. ui_find_views uses the same paging model. Omit snapshot_id to collect the current UI again; collection/filter options cannot change on a cached page. Default retention is 60 seconds, 4 captures and 8 MiB globally, with earlier eviction possible under the 2048-object device bound.

ui_automate reuses/attaches the verified app session and runs up to 20 actions in one MCP call (one Frida RPC for a semantic batch). Supported actions include tap, back, input_text, scroll, assert, wait_for and semantic accessibility operations. Targets accept a leased view_id, a unique identifier/class/text selector, or window coordinates in UIKit points.

{"device_id":"test","bundle_id":"dev.iosdecompiler.fixture","steps":[{"action":"tap","target":{"accessibility_identifier":"fixture.button"}},{"action":"assert","target":{"accessibility_identifier":"fixture.status","text":"Tapped"}}],"after_capture":"summary"}

Failed steps stop the batch; mutations are never retried automatically. Waiting is opt-in and bounded to 10 seconds per batch. Post-capture defaults to none; text and PNGs remain explicit.

Tap dispatches UIKit control events or accessibility activation. Back pops a navigation controller or dismisses a modal; scroll adjusts UIScrollView offsets; input sets UITextField/UITextView text and emits change events. These semantics cover app testing without additional device services. Physical gestures use the optional SSH input backend described below. See automation and value queries for examples and limits. Paged feeds accept scroll with pages: 1 or pages: -1; distance comes from the current scroll viewport on-device. Reuse its observed view ID and opt into animated: true/settle_ms when the app needs animation before checking state.

objc_query_instances combines attach/reuse, bounded heap discovery and concrete ivar values. Select fields to keep results small:

{"device_id":"test","bundle_id":"dev.iosdecompiler.fixture","class_name":"DecompilerFixtureViewController","field_names":["_tapCount","_button"],"max_instances":5,"limit":5}

Use object_ids instead of class_name to read already leased objects, including different classes, without scanning the heap. Pages use the returned snapshot_id and never revisit the device. Default reads cover at most 8 ivars, 20 instances, 10 rows per page, bounded Foundation values and redacted sensitive fields. This Objective-C query invokes no arbitrary getter or description. Pure Swift and explicitly typed native operations use their own advanced tools below.

ui_capture defaults to backend="auto": UIKit summaries gain public SwiftUI accessibility-container nodes when exposed. backend="accessibility" selects that semantic tree. Its leased IDs share the existing snapshot cache and work with ui_automate tap or action="accessibility" operations: activate, increment, decrement, escape and scroll. Text remains opt-in. This does not reconstruct the SwiftUI internal view/state graph or enable VoiceOver globally.

Advanced tools remain behind platform_tool; focused discovery stays at 20 tools.

{"name":"swift_runtime","arguments":{"device_id":"test","bundle_id":"dev.fixture","action":"types","images":["MyFramework"],"kind":"classes","query":"Player","limit":5}}

The official frida-swift-bridge 3.0.2 is experimental. Vesper patches its import-time whole-process scan into lazy image-scoped discovery with a type budget (default 1000); default scope is the main executable and libswiftCore. Capability queries do not load the bridge. It reuses the native Frida session, caches metadata, and can reset independently. inspect(kind="class", address=…, field_names=[…]) executes only explicitly requested getters. Retained class handles use Swift ARC, expire with the configured object TTL, and are capped at 128; active calls protect inputs. Struct/enum storage is borrowed; scalar values/tag summaries need no deep copy or retained handle. Generic, async, stripped and floating-point signatures can remain unsupported. There is no automatic pure-Swift heap scan.

Native calls require a known C ABI signature; C++ member calls require an explicit this pointer. Read selected fields with known offsets from IDA/debugger evidence. Return pointers and temporary C strings are borrowed. Do not guess an ABI.

swipe, long_press and gesture (1–5 finger paths for pinch/rotation/custom motion) route through optional ZXTouch over an SSH loopback channel. Set input_backend="touch" for physical tap/text/scroll. Raw coordinates use oriented screen UIKit points and need no Frida attach when no observation is requested. Screen size/scale are cached per SSH connection; orientation stays fresh. Cancellation attempts finger-up cleanup. A service acknowledgement confirms processing, not application behavior. The default UIKit/accessibility route needs no new service. ZXTouch is external, not bundled; input_setup returns actionable installation hints when unavailable. No VNC is required. See the on-demand Swift/native/input skill guide.

Opt in to each system target in your device configuration. Use the exact executable path reported by Frida on that device; these paths are examples:

[devices.test.allowed_processes.bluetoothd]
executable = "/usr/sbin/bluetoothd"

[devices.test.allowed_processes.sharingd]
executable = "/usr/libexec/sharingd"

The advanced process_list tool reports fresh PIDs, paths and path_verified for configured names. process_attach rediscovers the PID and requires a unique exact path match, then verifies the injected PID and main executable before opening the session. It never launches or restarts a service. SSH /rootfs mappings are separate from process identity and are not rewritten.

{"name":"process_attach","arguments":{"device_id":"test","process_name":"bluetoothd"}}

Pass the returned session_id to native analysis and logs; omit bundle_id:

{"name":"runtime_native","arguments":{"device_id":"test","session_id":"process_SESSION_ID","action":"modules","query":"bluetooth","limit":3}}

runtime_native also supports scoped symbols, selected inspect/read, and rate/event-limited trace on system sessions. Stop a trace with untrace or frida_unhook; read events with frida_get_events. Use app_logs with the same session, sources=["os_log"], a small limit, then action="stop". frida_list_hooks includes native hooks; frida_detach removes hooks and detaches. Allowlist changes are checked on subsequent RPCs/events. System operations run outside the main queue and require neither UIKit nor an Objective-C identity. Writes, native calls and UI/Swift operations remain unavailable on system sessions.

Start by attaching bluetoothd, inspecting identity.executable and a small module page, then detaching. Injection availability depends on the device's permissions; an attach failure reports its phase. Process identity fields use the official Frida Process API.

Open an allowlisted app on the unlocked phone. Call device_connect and device_list_apps with {"device_id":"test"}, then call frida_attach:

{
  "device_id": "test",
  "bundle_id": "dev.iosdecompiler.fixture"
}

Use the returned session_id in ui_snapshot:

{
  "device_id": "test",
  "session_id": "process_SESSION_ID",
  "depth": 8,
  "max_nodes": 500,
  "limit": 100,
  "include_text": false
}

Replace placeholder IDs with observed values. Use ui_inspect_view to inspect a returned view's controller and target-action mappings. Request text explicitly with include_text: true when needed. Call frida_detach when finished. Object IDs remain valid while a live snapshot or operation lease references them. They can expire or be evicted; reacquire them with a live capture. Session IDs also become invalid after idle reclamation, detach, process exit or restart.

After attaching, call ui_screenshot:

{
  "device_id": "test",
  "session_id": "process_SESSION_ID"
}

The tool renders the attached application's UIKit window to PNG, using the key window by default. To select another window, pass a window_id returned by ui_get_windows. It returns a native MCP image block plus structured metadata including mime_type, local_path, bytes and window_id. MCP clients with image support can display the image directly.

Images are saved under <state_dir>/screenshots/. Capture is limited to the selected application's renderable window; system overlays and some protected or GPU-rendered content may be absent.

The UIKit fixture provides a button with accessibility identifier fixture.button and an Objective-C action. Build, install and launch the fixture using that guide, then call develop_tweak:

{
  "device_id": "test",
  "bundle_id": "dev.iosdecompiler.fixture",
  "accessibility_identifier": "fixture.button",
  "replacement_title": "Hello vesper"
}

The workflow inspects the button, validates an observed controller lifecycle method, generates an editable Logos project and builds a package. It requires a unique matching UIButton and a supported method signature. Jobs run in the background by default; retrieve the report using job_status and the returned job_id.

At the approval stage, an operator reviews the exact operation in a separate terminal. Use the report's approval ID and the server's absolute state directory:

.venv/bin/vesper-approve --state-dir /absolute/path/to/vesper/.vesper approval_ID

Resume develop_tweak with workflow_id and the granted approval_id. After installation succeeds, restart through SSH with app_restart (already included by restart=true) and use Frida/ui_automate to exercise the requested change. Do not create backups, add local/smoke/long regression tests, or compare loaded modules, build markers, UUIDs or installed-file hashes. Inspect installation identity or crash reports only when diagnosing a relevant failure. Stop on the first important failure and fix it before continuing.

Complete supported build/install/restart/UI operations yourself. No VNC service is needed. After user acceptance, finish the authorized commit/push without adding more checks or repeatedly polling CI.

vesper/
├── src/vesper/
│   ├── server.py        # MCP stdio entry point
│   ├── config.py        # Validated configuration
│   ├── core/            # Sessions, state, approvals, jobs and evidence
│   ├── adapters/        # SSH, Frida, IDA, Mach-O and Theos integrations
│   ├── tools/           # Structured MCP tools
│   ├── workflows/       # Tweak development and deployment
│   └── assets/          # Bundled Frida agent and license notices
├── agents/frida/        # Device-side TypeScript source
├── examples/            # Configuration, UIKit fixture and demo
├── scripts/             # Build and documentation utilities
├── tests/
├── docs/
├── pyproject.toml
└── uv.lock

For app/tweak work, use this default flow:

Locate the issue → smallest change → build → install → focused device check → authorized commit/push.

Do not create backups or add local test suites, smoke tests, long regressions, sampling campaigns or test reports to that flow. Reuse what has already been read; inspect relevant differences instead of rereading entire files. Read only the target controls and capture an image when reproducing a visual issue, checking its fix, or diagnosing an exception. Do not continue checking after acceptance without a new change, failure or user request.

For Vesper code changes, use only the smallest relevant existing test. The lightweight CI builds the package and runs the deployment/restart tests once.

After changing Python source, update the installed MCP directly:

uv sync --frozen --all-extras --no-editable --reinstall-package vesper

Rebuild the bundled agent with npm ci --ignore-scripts && npm run build only when its TypeScript source changes. Generate tool documentation with PYTHONPATH=src .venv/bin/python scripts/document_tools.py.

• Device and application allowlists constrain targets. Routine inspection is the default; installation, app restarts and other disruptive operations require local approval.
• Approvals bind exact operation arguments and artifacts, expire, and can be consumed once. MCP clients cannot grant their own approvals.
• Tools expose bounded operations with structured errors, audit records and cleanup. Arbitrary shell commands and injected JavaScript are not exposed through MCP.
• UIKit operations, object retain/release and PNG encoding run on the main queue. Object leases expire or are evicted within the existing 2048-object bound; shared and executing operation references remain protected.
• UIKit/Objective-C, exposed SwiftUI accessibility, scoped experimental Swift and explicit native ABI operations use separate backends. Complete Swift metadata/SwiftUI reconstruction and all generic/async signatures are not guaranteed.
• Runtime/static address correlation requires matching image identity, architecture and segment evidence. Inferred types and pseudocode are labeled accordingly.

Keep local configs, keys, state databases, traces and screenshots private. See security and recovery for approval and rollback behavior.

Issues and pull requests are welcome. Include the relevant host, iOS, jailbreak, Frida and toolchain versions when reporting a problem, with sensitive data removed. Keep changes focused, use minimal relevant checks, and update both READMEs when changing user-facing behavior. Add new tools to the generated reference.

• LINUX DO

Original vesper code is available under the MIT License. Bundled code and dependencies retain their own licenses; see third-party notices.

Built with Frida, Theos and the MCP Python SDK, with inspection concepts inspired by FLEX.