Agents
Set up an optional AI agent (Claude Code, Codex, Gemini or OpenCode) to write and change your plugin's code from a chat.
An agent writes and changes your code from a chat. It's optional: Studio's editor, preview and builds all work without one.
An agent uses your own plan or API key with its provider, under that provider's terms. Your prompts and the project files the agent works on go to that provider; Freqlab never receives your sign-in, prompts or chats, and doesn't train on them. Agents can change or delete files, and you're responsible for what they do.
Pick an agentLink to this section
| Agent | Sign in with | Good for |
|---|---|---|
| Claude Code | A Claude subscription or API key | Working well out of the box |
| Codex | A ChatGPT plan or OpenAI API key | If you already pay for ChatGPT |
| Gemini | A Gemini API key, or Google for Workspace accounts | Google's models |
| OpenCode | Many providers, or models on your own computer | Mixing vendors, or running models locally |
Set one upLink to this section
- Open Settings, Agents, and find the agent under Your Agents.
- Click the card's install button. Codex and Gemini need Node.js; the card offers Install Node.js if it's missing.
- Click the card's sign-in button and finish in the terminal window that opens. Codex can also take an API key instead.
- Pick a Model, and click Set Default.
Already installed and signed in to an agent yourself? Turn on Manage this agent myself to skip Studio's checks.
Local models with OpenCodeLink to this section
On the OpenCode card, click Add Local Provider. Pick LM Studio, Ollama or vLLM (or enter a base URL), click Test Connection, check the models to use, and click Add Provider. Pick a model that supports tool calling: the agent reads and edits files with tools.
How the agent worksLink to this section
In Settings, Agents, under Behavior:
| Setting | Options |
|---|---|
| Permissions | Plan only: suggests without changing anything (not available with OpenCode). Auto-accept (the default): edits files and runs commands without asking (Gemini can't run commands here; use Full access for that). Full access: skips every check |
| Working style | Direct, Balanced or Thorough |
| Clear context automatically | Starts a fresh context when it's 80% full (adjustable), so the agent doesn't slow down. Your chat history stays |
| Custom instructions | Your own standing instructions for every chat |
Studio doesn't ask before each edit or command: the permission setting decides. OpenCode's Auto-accept, and Codex's on Windows, are the same as Full access.
To use a different agent or model for one project, open the project's ⋮ menu, then Agent & Model.
What the agent knowsLink to this section
Studio briefs the agent on every message, so it works the way Studio does without you explaining it:
- Your project: its framework and version, effect or instrument, and UI, plus your Working style and Custom instructions. It also follows your project's own
AGENTS.md,CLAUDE.mdorGEMINI.md. - Your Library: every guide and recipe, your own included. It checks for a match before it implements a feature or any DSP, and follows the tested approach over its own defaults.
- Audio safety: the rules that keep plugins from glitching or blowing up, such as no allocating or locking while processing audio.
- What Studio measured and caught: ask about CPU or a slowdown and it reads Studio's measurements, and can record your playing Profile build with Tracy. Report clicks, NaNs or silence and it reads what your probes caught; once probes are set up, it can add more to narrow it down.
- How Studio works: Studio does the building, so the agent writes and checks code but never bundles, installs or signs plugins, and tells you when it's ready to build. Studio saves a version of every turn, so the agent doesn't commit or change branches unless you ask.
None of this goes into your code or your git history.