MCP bridge between an AI agent and Stable Diffusion WebUI Forge Neo
Server has 12 tools with mostly complete descriptions (194-char baseline met for most). Naming follows verb_noun convention well (capabilities, model_profile, loras, models, generate, interrupt, skip, progress). Input schemas are present and typed for all tools. However, several critical gaps: (1) output schemas are not documented, responses are inferred from code but not formally specified in tool definitions; (2) error handling lacks recovery guidance, tools return {ok: false, error: string} but do not suggest next steps; (3) some parameter descriptions are vague (e.g., 'confirm' in prompt_dialect lacks format/range guidance); (4) no tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels; (5) generate tool has 12 parameters but no guidance on mutual exclusivity or defaults. Composition is sound, tools are single-purpose and chain well (loras → lora_info → generate). Security: no secrets in params (good), but no permission gates or audit trail declarations.
Report what this Forge instance offers: routes, counts, and which metadata sources are available. Call this first in a session.
Generate images from a text prompt using the currently loaded checkpoint.
Interrupt the currently running generation.
Full detail for one LoRA, including description, tags, past usage and a ready-to-paste prompt fragment with its trigger words.
Search available LoRAs by name, title, tags, trigger words or description. Only call this when the request actually calls for one (a named style, character, or concept) — most generations need no LoRA at all. `kind` can be "content" or "accelerator"; accelerators change the sampling regime rather than the image, so adopting one means adjusting steps and CFG together.
Describe the currently loaded checkpoint: architecture preset, whether it behaves as a turbo/distilled model, the sampling parameters that actually worked before, the expected prompt dialect, and whether its VAE and text encoder modules exist. Call before writing a prompt for an unfamiliar model.
Output schemas not documented. Tools return structured dicts but no formal schema specification in tool definitions. LLMs cannot plan downstream calls or extract fields reliably.
Error responses lack recovery guidance. Tools return {ok: false, error: string} but do not suggest next steps (e.g., 'Try search_loras() with a different query'). Agents cannot self-correct.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools declare risk levels in comments but not in schema. LLMs cannot infer safety properties from definitions alone.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |
List or load checkpoints. action: "list" | "load" | "refresh". Loading swaps the model for the whole instance, including any human using the web UI at the same time, and takes several seconds — only do it when the operator asked for that model. When the target belongs to a different architecture, its preset, VAE and text encoder are switched with it, since Forge would otherwise load it against whatever modules are selected now. The architecture is inferred from two signals and only acted on when they agree; pass `preset` to state it outright.
Check the VAE and text encoders loaded for an architecture against what it actually needs, and list installed files that could fill any gap. Defaults to the active preset. Worth calling after switching architecture or when output looks wrong for no obvious reason: Forge records the last selection made under a preset, so loading a checkpoint while another preset was active can leave the wrong modules attached. Where the reference does not state a VAE, it says so instead of guessing — a wrong VAE degrades output without raising an error.
Find, and optionally fetch, a VAE or text encoder the architecture needs. Called with no arguments it lists what the active preset is missing and where each file comes from, downloading nothing. Downloading requires both a `label` naming one entry and `confirm=True`, and the operator has to agree first: these are multi-gigabyte files written into their models folder, often across a network share. Links come from the Forge Classic wiki's Download Models page. Where several builds exist — bf16, fp8_scaled, gguf — they are all offered, because which to take depends on the operator's hardware, not on a default worth hiding.
Check the progress of the current generation.
How the loaded checkpoint expects to be prompted, with its quality tags. Returns the dialect (pony / illustrious / animagine / anima / sd15 / sdxl_base / natural), the quality prefix and negative baseline it needs, and where that conclusion came from. Quality tags are not decoration: an Illustrious prompt without them degrades, and a Flux prompt with them degrades too. When the dialect comes back unknown — `xl` covers Pony, Illustrious and stock SDXL, which share tensors and preset — ask the operator, then call again with `confirm` set to their answer. It is cached by file hash and never asked again.
Skip to the next image in the current batch generation.
generate tool has 12 parameters with no guidance on mutual exclusivity (init_image + denoising_strength only valid together), defaults, or constraints. Parameter descriptions lack format/range info.
prompt_dialect 'confirm' parameter lacks format guidance. Description does not specify valid values, length, or case sensitivity. LLMs may pass invalid dialects.