models.md

Models

Onde Code is built for model work on Onde Inference: writing and running the code that fine-tunes, manages and deploys your models. It does that with the same tools it uses for any repo (read and edit files, run commands) plus the Onde CLI, onde.

The Onde MCP server

onde --mcp runs the Onde CLI as a Model Context Protocol server over stdio. It exposes account and model-catalog operations as tools that return structured JSON:

Tool	What it does
login, me	Sign in, and show who you are signed in as
apps_list, app_create, app_rename	Manage your Onde Inference apps
models_list	List the models in the catalog
model_register	Register a fine-tuned GGUF model that is already uploaded to Hugging Face
model_assign	Assign a model to an app
hf_search	Search Hugging Face for base models

The tools run non-interactively and reuse the token from an onde sign-in. Install the CLI first, for example npm install -g @ondeinference/cli, and sign in with onde --tui.

Any ACP client can hand it to the agent in mcpServers, and Onde Code connects it for the session. In Zed, add it under context_servers in settings.json, or pass it to the agent through whatever MCP configuration your editor uses:

{
  "context_servers": {
    "onde": {
      "command": "onde",
      "args": ["--mcp"]
    }
  }
}

The tools then appear to the model as mcp__onde__models_list, mcp__onde__model_assign and so on.

Fine-tuning

Fine-tuning itself is part of the onde CLI, not of the MCP server: a local LoRA pipeline for Qwen2, Qwen2.5 and Qwen3 models that merges the adapter and exports GGUF. Onde Code can drive it with run_command like any other tool, and ask before it does. See the onde-cli README for the pipeline.

A typical session, from a repo with your training data:

1. Ask the agent to find a base model (hf_search) and set up the fine-tune.
2. It writes the config and runs the pipeline through onde, after you approve the commands.
3. Once the GGUF is on Hugging Face, it registers it (model_register) and assigns it to your app (model_assign).

Built in

Shipping the Onde tools inside the agent, with no MCP setup, is tracked in kelvin-soen/OndeCode#27. Until then, the MCP server above works today.