Loading ToolBench

Reading the current corpus…

modeldesk MCP server · D · ToolBench

← Browse

modeldesk

github.com/gao-shu/modeldesk

ModelDesk MCP server (stdio) — list_models / run_* over the shared run core

2026-07-28+STDIO7 tools

55/100

Definition Quality68

Protocol Readiness50

Supportability38

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence

Current-spec patterns detected

Summary

ModelDesk MCP exposes 7 tools with complete JSON schemas and descriptions. Naming follows verb_noun convention (list_*, run_*, cancel_*). However, descriptions lack LLM-optimization guidance (e.g., 'run_text' doesn't explain when to use it vs alternatives), parameter descriptions are minimal, and output schemas are undocumented. Error handling and recovery guidance are absent. No tool annotations (readOnlyHint/destructiveHint). The server is functional but falls short of production-grade quality.

Tools (7)

cancel_runreversiblesource verified73/100

Cancel an in-flight run started by this MCP process (same process only).

list_active_runsread onlysource verified63/100

List in-flight runs started by this MCP process (for cancel_run).

list_modelsread onlysource verified75/100

List ModelDesk-registered models for agent tools (text/image/video/audio). Same local DB as the Web UI when Desk is running (follows :3300 dataDir).

run_audioread onlyauthsource verified65/100

Run an audio model. modelId: registry UUID or unique config name. Optional params mirror UI.

run_imageread onlyauthsource verified65/100

Run an image model. modelId: registry UUID or unique config name. Optional params mirror UI.

run_textread onlyauthsource verified72/100

Run a text (chat) model. modelId: registry UUID or unique config name from list_models.

run_videoread onlyauth

Issues (5)

HIGHtool ↗

Output schemas undocumented. LLMs cannot predict what fields run_text, run_image, run_video, run_audio return, forcing them to guess downstream field names and risking failed chaining.

run_textrun_imagerun_videorun_audio

HIGHrecovery-guide ↗

No error recovery guidance. Descriptions do not explain what to do if a model run fails, times out, or returns invalid output. LLMs have no fallback strategy.

run_textrun_imagerun_videorun_audio

HIGHconstrained-input ↗

run_image, run_video, run_audio accept 'params' as additionalProperties:true with no schema. LLMs cannot discover valid keys (size, ratio, quality, reference_images) and will hallucinate invalid parameter names.

run_imagerun_videorun_audio

MEDIUMtool ↗

No tool annotations. Tools lack readOnlyHint/destructiveHint/idempotentHint, preventing LLMs from reasoning about safety and retry semantics. cancel_run is reversible but unmarked.

Recommendations

Document output schemas for all run_* tools. Example: run_text returns {runId: string, status: 'completed'|'failed'|'timeout', output: string, tokens_used: number, error?: string}.
Add error recovery guidance to run_* descriptions: 'If the model times out, call list_active_runs to check status. If failed, try again with lower temperature or fewer maxTokens.'
Replace 'params' additionalProperties with explicit schema for run_image, run_video, run_audio. Document valid keys: size, ratio, quality, reference_images, etc.
Add tool annotations: mark list_* and run_* as readOnlyHint=true; mark cancel_run as destructiveHint=true (or reversibleHint if supported).
Expand list_active_runs description to 50-100 chars: 'List in-flight model runs started by this MCP process. Use to check status before canceling. Returns runId, model, status, startTime.'

Score history

Overall score trend

First recorded score · v2 rubric

55/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	D	55	2026-07-28+	v2