AI Agent

Tucan Studio has a chat sidebar that talks to an AI service of your choosing. It is not tied to one vendor: the studio ships clients for OpenAI and Claude (Anthropic), and an extension can add more providers -- the Google AI extension adds Gemini, for example.

The chat works on the open workspace, so the Chat entry is disabled until you open a folder.

Adding a service

AI Agent -> Add AI Service... asks for:

• the provider (OpenAI, Claude, or any provider an installed extension contributes),
• your API key for that provider,
• a Workspace ID -- Anthropic only, and only for identity-linked keys. A standard key leaves it blank.

Services are stored in config/ai-agents.json next to the studio binary, so they follow the install and are shared by every workspace. The key is kept as you typed it; treat that file as a secret.

AI Agent -> Manage AI Services... lists what you have configured, and lets you edit or remove each one.

Chatting

AI Agent -> Chat opens the sidebar. Type in the input field and press Enter.

Across the top of the panel:

• Service -- which configured service this session talks to. Switching it also switches the model list.
• Model -- the models that service offers.
• Session -- chat sessions are separate conversations, each with its own history, service and model. + starts a new one and the pen button renames the current one.

Sessions are saved, so they survive closing the panel and restarting the studio. A session that is mid-request keeps working when you switch away from it; the reply lands in the session it belongs to.

Agent mode

The Agent checkbox lets the model use tools instead of only answering in prose. What it can reach:

• list / type -- browse directories and read files
• write / edit -- create a file, or replace text inside one
• run_command -- run a shell command
• web_request -- fetch a URL
• tucan_help -- list and read these help topics, so it can answer questions about the studio itself
• aminet_search / aminet_recent / aminet_download -- search Aminet, read the recent feed, and download a package into the folder set in Settings -> Downloads
• new_session -- start a fresh session carrying a summary it writes
• add_command / edit_command / remove_command -- maintain the studio's command list (see below)
• get_skills / get_skill / add_skill -- reusable instruction files it can consult or write

Writing a file, and running a command, are confirmed by you in a dialog before they happen.

The open file

Above the input is a checkbox naming the file in the active tab. While it is ticked, that file's contents are attached to each message you send, which is what lets you say "explain this" with no further context. The name follows the active tab; the tick is your choice and stays where you put it. It is remembered per session, so one session can work with the file attached while another does not.

The command list

The command list tells the agent which shell commands this system actually has. It is what stops a model reaching for Linux commands on AmigaOS: ask it to add one and it records the command with a short description, so later sessions use List rather than ls.

The list itself, and whether it is a hint or a hard allow-list, are in Settings -> AI Agent. Off means the agent may run only what is named there.

Instructions and context

Settings -> AI Agent also holds:

• Instructions given to the agent in every workspace, on top of the studio's own system prompt.
• The point at which older turns are compacted into a summary, as a percentage of the model's context window (default 75).

Slash commands

Typed in the chat input:

Provider-hosted tools

Some providers can run their own tools on their servers -- web search and code execution. The two toggles are in Settings -> AI Agent; with one on, the request asks the provider to use it and folds the results into the answer. They are independent of agent mode, which runs tools here on your machine, and a provider that does not offer them simply ignores the request.

Adding a provider

A provider is a small JavaScript file in an extension: it declares how to build a request for its API and how to read the reply, and the studio does the rest -- so it inherits agent mode, tool calling and the context readout without any further work. See docs/EXTENSIONS.md (aiProviders) and the google-ai extension as a worked example.