UI-Venus MCP
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| VENUS_MODEL | No | The model name to use. Defaults to UI-Venus-2-9B. | UI-Venus-2-9B |
| CUMCP_REMOTES | No | A comma-separated list of remote targets to register, e.g., windows-ssh:user@host | |
| VENUS_API_KEY | Yes | Your API key for the vision endpoint | |
| VENUS_BASE_URL | Yes | The base URL for the OpenAI-compatible vision endpoint, e.g., http://<gpu-host>:8300/v1 |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": true
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| computer_list_targetsA | Enumerate every device/platform this MCP server can currently serve (local desktops, android devices, ios simulators, browser), with availability and honest capability notes. |
| computer_get_targetB | Resolve one target and report its capabilities, permission state and platform details. |
| computer_get_stateA | Cheap structural state: screen info, frontmost app, windows, capabilities. No screenshot. |
| computer_inspectB | Full observation: screenshot + UI tree, optionally with a vision-model description focused on a topic. |
| computer_screenshotC | Capture the screen/window/display. Returns metadata + image content; optionally saves to a file. |
| computer_locateA | Find an element by natural language or structured descriptor. Structured (UIA/AX/AT-SPI/DOM) first, UI-Venus vision grounding as fallback; vision points are snapped back to structured elements when possible. Coordinates in results are screenshot pixel space; use with computer_action. |
| computer_actionB | Execute a single unified action (spec §17): click/double_click/right_click/long_press/move/drag/swipe/scroll/type/set_value/clear/press/hotkey/select/toggle/invoke/launch_app/terminate_app/focus/back/home/wait. Provide element (from computer_locate) or point. Records to the active recorder if recording. |
| computer_stepB | Observe + let the vision GUI expert decide the next action + execute it. Returns the decision and execution result. Useful for building your own loop in the calling agent. |
| computer_execute_taskA | Delegate a natural-language task to the autonomous GUI agent loop (observe → decide → locate → execute → verify → recover). mode=delegate|auto. Returns immediately with a taskId unless wait=true. Sensitive actions park in WAITING_CONFIRMATION with a confirmToken — re-call with the token to approve. |
| computer_get_taskB | Fetch a task record: state, steps, outcome, pending confirmation. |
| computer_cancel_taskB | Request cancellation of a running task; also cancels its pending confirmation. |
| computer_verifyB | Verify a goal against the current screen: structured UI-tree heuristics first, UI-Venus vision verification as fallback. Supports structured assertions (checked/exists/value/textVisible). |
| computer_record_startB | Start recording executed actions on a target session (semantic evidence: elements, hashes, apps). |
| computer_record_stopC | Stop recording and return the recorded entries. |
| computer_record_to_scriptC | Convert the current/last recording into a portable YAML automation script with semantic locators (never raw coordinates). |
| computer_run_scriptC | Execute a portable YAML automation DSL script (semantic locators, cross-platform) against a target. |
| computer_run_ui_testB | Run one or more YAML UI test cases; returns PASS/FAIL/SKIP/BLOCKED per case with step evidence. BLOCKED = environmental restriction (permissions/offline/unsupported), reported honestly. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 17 tools
Most tools have clearly distinct scopes (single action vs. one autonomous step vs. full task delegation; different observation depths; recording vs. script execution). Some boundaries among computer_action, computer_step, computer_execute_task, and computer_run_script could blur, but descriptions and parameter differences largely disambiguate them.
All tool names use consistent computer_ snake_case with clear verb/noun or record_* patterns. There is no mixed casing or conflicting verb style.
17 tools is slightly above the typical 3-15 sweet spot, but the breadth of GUI automation capabilities (targets, observation, action, tasks, recording, scripting, testing) justifies most tools. A few tools could potentially be consolidated but are reasonable for this domain.
The surface covers target discovery, observation, element location, action execution, autonomous task delegation, verification, cancellation, recording/scripting, and UI testing. There are no obvious dead ends for the stated GUI automation purpose.