Local GPU Imagegen
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| LOCAL_GPU_IMAGEGEN_STATE_DIR | No | Override for the OS user-state directory where trust state is stored. Trust is stored outside the repository under the OS user-state directory, overridable with this environment variable. | |
| LOCAL_GPU_IMAGEGEN_OUTPUT_DIR | No | Override for the output directory. Compatibility-tool files default to outputs/. High-level runs use the runs/<run_id>/ layout under that output root. | |
| LOCAL_GPU_IMAGEGEN_REALESRGAN_DIR | No | Configure the tool root for the optional anime-only Real-ESRGAN postprocess. The server accepts only realesrgan-ncnn-vulkan.exe plus one of the supported model pairs, realesrgan-x4plus-anime or realesr-animevideov3-x4, under that root. It accepts no arbitrary executable path or model name, and it does not download a binary or model. |
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 | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| local_gpu_imagegen_checkA | Check Python packages and CUDA readiness for local GPU image generation. |
| local_gpu_generate_imageB | Compatibility tool for WebUI or Diffusers image generation; not for ComfyUI routes and never a bypass for the confirmed high-level run workflow. |
| local_gpu_discover_modelsC | Plan or execute bounded local model discovery without loading model weights. |
| local_gpu_inspect_workflowA | Inspect one local ComfyUI API workflow and infer a safe ordinary txt2img binding without writing state. |
| local_gpu_register_workflowA | Recheck and immutably register one exact previously inspected ComfyUI workflow proposal. |
| local_gpu_set_model_trustB | Inspect, approve, or revoke one exact current local model identity. A trust mutation requires the exact confirmation previously returned for the same mutation boundary, displayed to the user, and repeated in a later user message. |
| local_gpu_recommend_modelsA | Recommend one exact confirmed-capability route and at most two alternatives. preferred_model_id must be an exact catalog ID. Display the selected route and start_run_boundary, then wait for later user confirmation before local_gpu_start_run. |
| local_gpu_list_profilesB | List registered visual-asset profiles and current local backend capabilities. |
| local_gpu_start_runA | Create a visual-asset run only after later user confirmation; copy start_run_boundary exactly from the previously displayed recommendation and add the remaining run fields. |
| local_gpu_get_runB | Get the current persisted state of a visual-asset run. |
| local_gpu_branch_runC | Create an immutable child revision run from one reviewed parent round. |
| local_gpu_prepare_maskC | Prepare an unconfirmed child-run inpaint mask and return its JPEG overlay. |
| local_gpu_confirm_maskC | Confirm an unchanged prepared mask after explicit overlay approval. |
| local_gpu_generate_roundC | Generate one root or immutable revision round and return an optional bounded JPEG preview. |
| local_gpu_record_reviewB | Record human or model review evidence for one generated round. |
| local_gpu_finalize_runB | Publish the selected reviewed round as the run's final PNG. |
| local_gpu_cleanup_runC | Remove run intermediates or a fully confirmed run directory. |
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
Each tool has a clearly defined role in the local GPU image generation pipeline, from environment checks to run finalization. The two generation tools (generate_image and generate_round) are distinct, with one serving as a compatibility fallback and the other as part of the core workflow.
All tools follow a strict 'local_gpu_<verb>_<noun>' convention using snake_case. The pattern is uniform and predictable, making it easy to deduce tool behavior from the name alone.
At 17 tools, the set is slightly above the typical 3-15 range, but the complexity of the domain (workflow inspection, trust management, run lifecycle, and mask handling) justifies the count. Each tool serves a distinct purpose, so the overhead is acceptable.
The toolset provides end-to-end coverage for managing local GPU image generation runs, including setup, registration, recommendation, execution, review, and cleanup. Minor gaps exist, such as no explicit list-runs tool and no update operation for run fields, but these are not critical given the immutable nature of runs.