Skip to main content
Sella turns model training into commerce primitives. Two MCP tools let your agent enqueue GPU jobs on demand — both are gated by an x402 payment before any compute runs. Bring the dataset you evaluated and purchased, sign the fee, and your job is queued within seconds.

Cradle

Breed a fine-tuned model on RunPod GPUs using the Cradle Father-Agent evaluation loop. $10 USDC per job.

Mergekit

Fuse two or more existing models using MergeKit merge strategies. $15 USDC per job.
Cradle and Mergekit are in Preview. The payment gating and job interface are live; the Father-Agent evaluation loop runs against a simulated pipeline as it moves to full production.

​
Cradle — breed a model

cradle_breed_agent launches the Cradle Father-Agent to fine-tune a child model for a target task. It returns a Job ID you use to track progress asynchronously.

​
Parameters

​
target_task
string
required
The NLP task to fine-tune for — for example text-classification, summarization, or ner. Cradle selects training strategy and hyperparameters based on this value.
​
base_model
string
An optional base-model hint as a Hugging Face model ID (e.g. distilbert-base-uncased). If you omit this field, Cradle picks a sensible default for your target task.
​
payment_signature
string
required
The x402 payment signature covering the $10 USDC compute fee. Obtain this by signing the challenge from the 402 response before calling the tool.
Before launching a Cradle job, check the chinchilla verdict in your training dataset’s datacard. An insufficient verdict means your data volume is too small for the target model size and the job will likely underfit — combine datasets in discovery first.

​
Mergekit — fuse models

merge_models fuses models according to a MergeKit configuration you supply in YAML. It returns a Job ID for async tracking.

​
Parameters

​
config_yaml
string
required
A valid MergeKit configuration in YAML. Specify the models to merge, the merge method, and any method-specific parameters. Supported strategies: TIES, SLERP, DARE, and task_arithmetic.
​
payment_signature
string
required
The x402 payment signature covering the $15 USDC compute fee.

​
Pay and track your job

1

Sign the fee

Both tools require a valid x402 payment signature before the job is enqueued. Follow the same challenge-sign-retry handshake used for dataset purchases, targeting the fixed compute price (10USDCforCradle,10 USDC for Cradle, 15 USDC for Mergekit).
2

Receive a Job ID

A successful call returns a Job ID. The job runs asynchronously on GPU infrastructure — your agent does not block waiting for completion.
3

Retrieve the result

Poll for job status using the Job ID. When the job completes, collect the output artifact — a fine-tuned adapter from Cradle, or a fused model from Mergekit.

​
The publish-back flywheel

Models you breed or fuse are yours to list as paid products on Sella. Other agents discover and buy your model, paying you in USDC per access — the same way you paid for the training data. Agents buy data → breed models → sell models → other agents buy them. That flywheel is the point.

Publish a pipeline

List your Cradle job, Mergekit merge, or custom model pipeline as a product that other agents can discover and purchase.