The workspace
The workspace
Chat, coding, and the live session — how you actually work with Anvil day to day.
The workspace is where you and Anvil get things done. Whether you start in the terminal or the browser, it's one live session: a conversation, the files and commands it touches, and a rail that shows what the agent knows and is doing.
Chat vs. Coding
Anvil opens on Chat — ask anything, and it answers, runs tools, and streams the result. When the work is building or changing a codebase, switch to Coding: same engine, but oriented around a project — file edits, diffs, running the code, and committing.
- Chat — a conversation that can reach for any tool (bash, file read/write, web search, image generation, and more).
- Coding — a project-scoped session: open a repo, let Anvil read and edit it, run it, and prepare commits.
Both run against whatever model you've selected, and both carry the same memory and permissions.
Picking a model, mid-session
The model selector at the top of the workspace sets the active provider and model. Switch it any time — Anvil swaps the provider, routing, and system identity together, so you're never half-changed between two providers.
/model gpt-5.6 # switch provider + model on the fly
/model # or pick from the list
Permissions: you decide what it can do
Every session runs under a permission mode you control — from read-only to full access. A destructive or outward-facing action asks first unless you've granted it. The rail shows the current mode at a glance.
[!NOTE] Permissions are per-session and explicit. Anvil won't take an irreversible or outward-facing action on your behalf without the mode or a confirmation that allows it.
The status rail
Down the side of the workspace, the rail is your situational awareness:
- Model — the active provider/model and effort.
- Permissions — the current mode.
- Memory — what Anvil has learned and is carrying into this turn.
- Mind map — a live, force-directed view of what it's thinking.
- Sessions — switch between conversations; search and filter them.
Related
- Memory — how the rail's memory works, and how you curate it.
- Command reference — every slash command, including the session and model commands used here.