hunyuan-workbench
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| HUNYUAN_WORKBENCH_ROOT | No | Path to the data directory. Overrides the default location (e.g. %LOCALAPPDATA%\hunyuan-workbench on Windows). Must be set consistently for all calls using the same task set. |
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": false
} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
| experimental | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| workbench_capabilitiesA | Observed feature catalog, inputs, tested status and current limitations. Excludes world generation. |
| workbench_prepareB | Validate/copy/deduplicate 3..8 views including front. Does not submit or consume quota. |
| workbench_prepare_featureC | Validate a feature specification and preserve local inputs. See capabilities and README for schemas. |
| workbench_listA | Read durable local jobs; no browser calls. |
| workbench_getA | Read one job, its input hashes and state history; no browser calls. |
| workbench_browser_startA | Start/reconnect the persistent local browser. Uses its existing official web login. |
| workbench_browser_runA | Upload and configure the full job in one local execution. submit=True generates once and starts local monitoring with automatic download; consumes web quota. |
| workbench_submitA | Submit an already prepared READY form exactly once. Use only with user authorization for this job; may consume web quota. |
| workbench_statusC | Read local state or enqueue a fresh check of the exactly bound web asset. |
| workbench_watchA | Configure persistent local task monitoring. Checks happen in the worker without repeated model calls. Stopping monitoring does not cancel generation. |
| workbench_watchesA | Read monitor results, errors, next check times and attention flags from disk. |
| workbench_downloadB | Select the matching completed asset, download via official UI, validate and archive. Never generates again. |
| workbench_assetsA | List currently loaded visible asset cards and thumbnail identities on this job page. Does not enumerate unloaded pages. |
| workbench_recoverC | Explicitly rebind an interrupted task only after reviewing its matching completed asset. Thumbnail must identify the one selected card. Never generates. |
| workbench_browser_inspectA | Read the visible job page or default page for diagnosing a changed UI. |
| workbench_browser_resultA | Retrieve one local request result. Read watches for generation progress; do not repeatedly inspect the page. |
| workbench_ingestB | Validate/archive a manually downloaded artifact for a confirmed successful job. |
| workbench_blender_scriptC | Create a separate background Blender import script with metres, normalized height and grounded origin. |
| workbench_doctorB | Inspect local runtime, worker heartbeat and persisted monitoring state. |
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 19 tools
Multiple read/state tools (list, get, status, watch, watches, browser_result, doctor) have overlapping 'read local state or monitoring results' purposes, and browser_run vs submit can both lead to generation. The descriptions do help distinguish them, but an agent still faces several near-boundary choices.
All tools share the workbench_ prefix and snake_case, which gives a readable family feel. However, the naming convention is mixed: some are verb_noun like browser_start and prepare_feature, while others are bare nouns or verbs like list, get, watches, assets, capabilities, and doctor.
At 19 tools, the server is in the borderline-heavy range. Most tools map to a real workflow stage, but several status/watch/result variants could likely be consolidated.
The server covers the core generate-monitor-download-archive lifecycle, including recovery, manual ingest, diagnostics, and basic Blender integration. Minor gaps exist around cancellation and cleanup/delete operations, but the intended workflow has no obvious dead ends.