Writing messages

The composer is the message box at the bottom of every chat. You write what you want there, choose which model answers, and send. It also takes files, images and voice, and it can turn a message into a Goal, Swarm or Plan mode request.

Send a message

Type your message and press Enter. Shift+Enter adds a new line. While the agent works, the send button turns into a stop button: click it to stop the current turn.

A new session starts on the welcome screen. Nothing is created until you send the first message. That message creates the session with your defaults (provider, model and permission mode) and goes to the agent.

Choose where the session works

Above the welcome composer sits the workspace chip. It shows the project folder the next session will use, "No project", or "Choose workspace" when nothing is set. Click it to:

• search and pick one of your projects;
• choose New project to pick a folder with the system folder picker;
• choose Don't work in a project for a chat with no project folder. These sessions work in a folder named agenc in your home folder, which the app creates when needed.

Your home folder itself cannot be a project. The app says "Choose a project folder other than your home directory." Opening a project in the sidebar also points the chip at that project, so the next session starts there.

Choose the model, effort and speed

The pill at the right of the composer shows the model. Click it for a menu with these rows:

• Model: switch between the models pinned for this provider.
• Effort: how hard the model thinks before it answers. Faster costs less, higher effort reasons longer. The levels a model offers come from its own documentation, from None and Minimal up to Extra high and Max. Extra high and Max are marked "Consumes usage limits faster".
• Speed: Standard or Fast ("Faster output at a higher price"). The row appears only for models whose provider sells a faster tier: OpenAI priority processing, Anthropic fast mode, and xAI priority processing for Grok 4.7 and 4.6. xAI charges twice the standard rate for Fast, and only on an API key: a Grok account signed in with X does not offer Fast.
• All models: opens the model picker.

Each session keeps its own model, effort and speed. On the welcome screen the pill changes your defaults for new sessions, the same ones as in Settings > Providers. If you switch the model of a session while a turn is running, the app says "(model) takes over when this turn finishes" and switches afterwards.

The model picker

All models, or the model picker shortcut (Ctrl/Cmd+M by default), opens a searchable list of every model the provider offers, with what each one can do: context size, whether it reasons in steps, whether it reads images. From there you can:

• search models, or click Switch provider to search providers and move to another one. Providers without a key say "Add its API key in Settings → Providers first";
• click Fetch list to read the provider's current model list;
• pin a model to the composer pill with Show in the composer chip. The pill holds up to 7 models per provider. When it is full, unpin one first;
• type a model id the list does not have and pick the row that offers to use it, tagged "custom id". This is offered only for providers whose model list is not authoritative.

Local models also show their quantization and whether their weights are loaded in memory. When the app has verified an encrypted connection to an AgenC hosted model, the pill shows a TLS badge ("Encrypted connection to the model server (TLS 1.3).").

Which providers and models appear depends on what you connected. See Providers and models.

Attach files, images and audio

There are three ways to attach something:

• the + button, then Files and folders, which opens a file picker;
• paste an image from the clipboard;
• drag files onto the window. You can drop them anywhere in the window, and the composer shows "Drop to attach". Images dragged out of a web page are fetched and attached too.

Attachments appear above the text. Remove one with its close button.

• Images go to the model as images when the model reads images. Otherwise the agent gets the file path.
• Other files go to the agent as file paths, which its file tools can open.
• An audio file is transcribed by Core before the message is sent, and the agent gets the text. Core prefers the session's provider when that is OpenAI or Gemini, and a local engine otherwise. The chip shows "transcribing…", then "voice note" with its length, and its tooltip names who transcribed it. If transcription fails, the chip says "transcription unavailable" and offers Retry transcription.

Limits: up to 8 attachments per message and 50 MB per file. Images are limited to 7 MB in total per message, and Z.AI accepts only images under 5 MiB each. An audio file must be 10 MB or smaller to be transcribed.

The + menu

The + button opens the Add menu:

• Files and folders: attach a file.
• Goal: "Implement it, prove it, have it reviewed".
• Swarm: "Split the work across agents".
• Plan mode: "Explore, ask, then plan before touching anything".
• Plugins: every installed and enabled plugin. Picking one inserts @plugin-name into your message so the agent uses it.

A mode you pick shows as an icon beside the permission chip. Goal, Swarm and Plan mode stay on for the following messages until you remove them. Goal cannot be combined with the other two. Starting a message with the word "plan" (or "planning", "planea", "planifica") also turns Plan mode on. The modes are explained in Goal, Swarm and Plan.

The permission chip beside the + button sets how much the agent may do without asking. See Permissions and approvals.

The @ and / menus

Type @ at the start or after a space, or / at the start of a line, and a menu opens with what you can call by name:

• installed plugins;
• skills, with where each skill comes from;
• with / only, plugin commands. Once you pick a command, the box shows a grey hint for the arguments it takes.

The menu ranks exact names first. Use the arrow keys and Enter, or click. In your sent message, known names show as the plugin or skill name. Core runs the skill or command; skills are explained in Plugins and skills and in the AgenC Core guide.

One command belongs to the app itself: /resolve. It records what a tool call did when the app could not tell after a crash. See Sessions and projects.

Messages you send while the agent works

You do not have to wait for a turn to end. A message sent while the agent is busy joins a queue above the composer, and queued messages go out in order as each turn ends. The queue is saved as you change it, so it survives a reload of the window.

For each queued message you can:

• drag the handle to change the order;
• click Steer to send it now. This stops the current turn;
• remove it from the queue;
• open More for Edit message, Open in side chat (starts a new session with that message) or Turn off queueing. With queueing off, the session refuses new messages while a turn runs.

When you stop a turn, the queue pauses and says "Paused after stop." Click Resume to send the rest. A queue also pauses when its first message did not reach Core, or may not have. If the app cannot save the queue it says "These messages will not survive a reload."

Run a shell command from an answer

When an answer contains a shell command in a code block, the block gets a Run button. Click it to run the command in the terminal beside the chat, in the session's folder. The output appears under the block, labeled "Running", "Output" or "Exit code" with the number. Show terminal opens the terminal tab. If the command waits for input, the block says "Waiting for input. Type in the terminal." This is how you run a command that needs a password or a key: the agent writes it, you run it and type the secret in the terminal, never in the chat.

Details and limits:

• Run appears on blocks tagged bash, sh, zsh, shell or console on macOS and Linux, and powershell, pwsh, ps1, cmd or bat on Windows. It appears once the turn has finished.
• A block that contains hidden characters cannot be run. The button says "Command contains hidden characters".
• If the shell is not installed, the button explains why and stays off.
• The terminal is described in Browser and terminal.

Voice

The microphone in the composer turns on Whisper, which listens, transcribes on your computer and sends each phrase after a pause. See Whisper voice input.