Claycable

List Models

List all fine-tuneable text/language models available across Together.ai, Fireworks, and OpenAI. Returns model ids, sizes, context lengths, and estimated training costs.

List Image Models

List all image fine-tuning models available across Replicate (Flux Dev LoRA, SDXL DreamBooth LoRA, SD 1.5 LoRA) and FAL.ai (Flux LoRA Fast, Fast SDXL LoRA). Returns model ids, providers, sizes, and descriptions.

Estimate Image Cost

Estimate the cost of an image fine-tuning job on Replicate or FAL.ai based on provider, model, and training steps.

Scan Duplicates

Scan a set of Q&A training pairs for near-duplicates using TF-IDF cosine similarity. Returns clusters of similar pairs above the threshold. Use this before training to clean datasets and save token costs.

Estimate Cost

Estimate inference cost for a prompt across multiple models. Returns cost per request in USD sorted cheapest first. No API calls made.

Compare Models

Run the same prompt against multiple models in parallel and compare responses, latency, and cost. Use this to help users pick the best model for their use case.

Import Adapter

Import a fine-tuned adapter from a HuggingFace repository into the Vault. Fetches metadata, LoRA config, and base model info. Does not download weights.

Run Inference

Run a single inference call against any OpenAI-compatible model. Returns the response text.

Get Server Config

Check which provider API keys are configured on the server (Together.ai, NVIDIA). Returns booleans — never exposes key values.

List Vault Prompts

List the latest version (head) of every prompt in the workspace Prompt Vault (slug, title, content, tags, version).

Get Vault Prompt History

List all saved versions of one vault prompt slug, newest first.

Patch Vault Prompt

Update the current head version in place (autosave). Omit fields you do not change. At least one of title, content, or tags should be provided.

Publish Vault Prompt Version

Create a new numbered version for a slug (v1 for new slug, or increment). Use for intentional snapshots after edits.

Restore Vault Prompt Forward

Copy an older version forward as the newest snapshot (like “restore” in the Vault UI).

Datasets.create

Create a new empty training dataset in the workspace. Returns the dataset id.

Datasets.list

List datasets in the workspace (excludes soft-deleted). Returns metadata + pair count.

Datasets.get

Get a single dataset by id — metadata and pair count (not the pairs themselves).

Datasets.update

Patch a dataset (name, description, system_prompt, target_model, tags). Provide only the fields you want to change.

Datasets.delete

Soft-delete a dataset (and all its pairs). Irreversible via MCP.

Datasets.pairs.append

Bulk-append user/assistant training pairs to a dataset. Validates non-empty strings. Returns inserted count.

Datasets.pairs.list

List training pairs in a dataset. Optionally filter by category. Paginated.

Datasets.pairs.update

Patch a single training pair (user, assistant, category, metadata). Provide only changed fields.

Datasets.pairs.delete

Permanently delete a single training pair.

Datasets.export Jsonl

Export a dataset as JSONL ready for Fireworks fine-tuning. Format: chatml (default) | openai | fireworks — all produce the same OpenAI-compat messages format accepted by fireworks_finetune.py.