Provides Android automation capabilities, including reading the screen, tapping and typing, remembering UI elements as named anchors, verifying screen state with checkpoints, and running resumable multi-step routines on connected Android devices, emulators, or BlueStacks.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@digitomancy-mcp run my morning routine on the Pixel and resume if it fails"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Divination by fingers. Android automation for people and agents: read the screen, tap and type, remember buttons as named anchors, check where you are with checkpoints, and run multi-step routines that resume after a failure. Works with a USB phone, an emulator or BlueStacks, from a CLI, from Python, or from Claude as an MCP server and plugin.
Status: early but working. Tested on a Pixel 9 (Android 16) plus unit tests against a fake device. The Tasker bridge has unit tests but hasn't yet been verified end to end on a device.
Tool | Good for | Why it isn't the whole answer |
Ad-hoc agent poking via the accessibility tree | No memory: every session finds the same buttons again | |
A full LLM agent framework | Heavy and opinionated if you already have agents | |
Fast Android-only driver, no Appium server | Digitomancy builds on this | |
Appium | Cross-platform QA | Needs a server; overkill for Android-only |
Mirroring the screen | Use |
What Digitomancy adds: named devices, anchor maps (selector first, normalized coordinates as fallback), checkpoints with failure snapshots, resumable routines, and a Tasker bridge. What an agent finds once gets written down; the next run looks it up.
Related MCP server: duet
As a Claude plugin: build it with scripts/build_plugin.sh and install dist/digitomancy.plugin. You need adb and uv on your PATH. See plugin/README.md.
As a CLI or library:
git clone https://github.com/MadnessEngineering/digitomancy-plugin && cd digitomancy-plugin
uv sync --group dev # add --extra vision for image-template checkpoints
uv run dgm doctor # tools, workspace, which devices are onlineAs an MCP server in any MCP client:
{"mcpServers": {"digitomancy": {"command": "uv", "args": ["run", "--project", "/path/to/digitomancy-plugin", "digitomancy-mcp"]}}}Phone: enable Developer options (tap Build number 7 times), turn on USB debugging, plug in, and accept the prompt. On first connect uiautomator2 installs its helper on the device.
BlueStacks: go to Settings → Advanced and enable ADB, then uv run dgm -d bluestacks connect. Check the port it shows; edit devices.yaml to match.
Your devices, maps and routines live in a workspace: $DGM_HOME, or ~/.digitomancy by default, created on first run with starter maps and routines. Existing files are never overwritten.
devices.yaml: names → serials. Pick one with -d NAME, $DGM_DEVICE or default:. Unknown names are used as raw serials.
maps/<app>.yaml: anchors (things to tap), screens (checkpoint specs) and hazards (things never to tap). Refer to them as map:name.
routines/*.py: routines built from stages. A stage can require a screen, and a failed run can resume_from that stage.
.runs/: evidence from failed checkpoints (screenshot, UI dump, foreground activities).
catalog/ (optional, from dgm catalog): a redacted index of your Tasker projects and tasks. hazards_manual.txt and hazard_patterns.txt there tell the Tasker bridge which tasks it must never fire.
Coordinates: a float in [0,1] is a fraction of the screen, and an int is a pixel. Maps store fractions so they work at any resolution. They still depend on orientation.
Checkpoints: a screen spec combines conditions with all or any: selector, gone, app, activity, anchor, image. A failed expect saves a snapshot to .runs/.
Tasker: dgm tasker "Task Name" -p par1=x fires a task through am broadcast. It needs a one-time profile on the phone; see docs/TASKER.md.
dgm doctor | connect | look [--all] | tap map:anchor|x,y | check map:screen
dgm routines | run NAME [-p k=v] [--resume-from STAGE]
dgm tasker TASK [-p k=v] [-w SEC] | catalog [SOURCES…]
dgm mirror [-- scrcpy args] | mcpAccessibility suppression. While uiautomator2's on-device server runs, Android suppresses every accessibility service: Tasker, AutoInput, Join, screen readers. The CLI stops the server when each command exits; the MCP server stops it after 60s idle or on release.
scrcpy --new-display. Apps on virtual displays don't count as foreground. Detection reads Display #0 only.
Launching apps. Device.launch() resolves the default launcher activity. uiautomator2's app_start can pick a random one, and some apps (Tasker) run a task from their second launcher activity.
uv run pytest -q # no device needed: tests use a fake device
./scripts/build_plugin.sh # dist/digitomancy/ and dist/digitomancy.plugin
claude plugin validate dist/digitomancyMIT. Copyright (c) 2026 Madness Engineering. The Tasker action code table in the plugin comes from Taskomater/Tasker-XML-Info (MIT).
Evaluate a mapped screen 'map:screen' (optionally waiting). Returns ok + per-condition detail.
Inputs: screen (required), device, timeout. Returns: object
Registered devices (devices.yaml), the default, and which serials adb sees right now.
Inputs: no named parameters. Returns: object
Press a key: home, back, enter, recent, power, volume_up, volume_down, delete…
Inputs: name (required), device. Returns: result
Start an app by package name.
Inputs: package (required), device, stop_first. Returns: result
All maps with their anchors and screens.
Inputs: no named parameters. Returns: object
Known routines with params and stage names (stages are resume points).
Inputs: no named parameters. Returns: result
Current app + compact UI node list. Lines: #idx class 'text' id=res @(x,y normalized) [tap].
Inputs: limit, device, everything. Returns: result
Stop uiautomator on the device so Tasker/AutoInput accessibility comes back.
Inputs: device. Returns: result
Run a routine. On failure returns the failed stage, snapshot dir, and a resume hint.
Inputs: name (required), device, params, resume_from. Returns: object
Remember a tap target. From a look node (stores selector + point) or raw normalized x,y.
Inputs: name (required), map_name (required), x, y, +3 more. Returns: result
PNG of the screen (downscaled to save tokens). Use when look can't see it (games, canvases).
Inputs: scale, device.
Raw adb shell escape hatch.
Inputs: cmd (required), device. Returns: result
Swipe between two points (fractions or pixels). Scroll down = swipe(0.5,0.75,0.5,0.25).
Inputs: x1 (required), x2 (required), y1 (required), y2 (required), +2 more. Returns: result
Tap. Floats in [0,1] are screen fractions (preferred); otherwise pixels.
Inputs: x (required), y (required), device. Returns: result
Tap a mapped anchor 'map:name' (selector first, saved point as fallback).
Inputs: ref (required), device, timeout. Returns: result
Tap node #index from the last look on this device.
Inputs: index (required), device. Returns: result
Search the Tasker catalog (built by dgm catalog) by task/profile name substring.
Returns matching tasks with their actions and calls, plus hazard flags. Never fire a task marked hazard; tasker_run refuses them anyway.
Inputs: limit, query, project. Returns: object
Fire a Tasker task via broadcast (needs the DGM Dispatch profile, docs/TASKER.md).
Inputs: task (required), device, params, wait_result. Returns: object
Type into the focused field.
Inputs: text (required), clear, device. Returns: result
Where devices.yaml, maps/ and routines/ live (created and seeded if missing). Edit devices.yaml there.
Inputs: no named parameters. Returns: object
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Scored across 20 tools
Most tools have clearly distinct purposes, but tap, tap_node, and tap_anchor overlap as three tap variants, and tasker_run versus run_routine could be confused without careful reading. The descriptions generally clarify when to use each tool, so ambiguity is limited.
There is a mix of snake_case multiword names and bare nouns/single verbs such as shell, workspace, look, tap, and swipe. The convention is readable but not consistent across the set.
20 tools is on the heavy side for a device-automation server, though most correspond to real primitives rather than duplicates. It sits in the borderline-heavy 16–25 range.
The set covers device inspection, interaction, maps/anchors, routines, Tasker, and raw adb shell. It lacks structured CRUD for maps and routines and some UI actions like long-press, but the shell escape hatch prevents hard dead ends.
Drive cloud Android phones on US mobile IPs. Read the screen, tap, type, install apps and run tasks.
Cloud Android phones for AI agents: read the screen, tap, type, install APKs, park and resume.
Give your AI agent a real cloud Android phone: push APKs, drive it, get screenshots and logs.
Let AI agents control Android and iOS devices on Mobilerun Cloud: screen, input, apps, workflows.