Agent Providers

Overview

Settings → Agent Providers lets you connect agent runtimes like Cursor and Claude Code to run code-aware models on your workspace. Unlike cloud providers, agent-provider models don't just answer — they can read and edit files in a folder you choose.

Each provider is a row in an accordion. The key icon turns green when the provider is connected (an API key is saved for Cursor; the switch is enabled for Claude Code).

Agent providers require Pro or higher. On the free plan the page is locked with an Upgrade button.

---

Cursor

Cursor connects with an API key:

1. Expand Cursor and click Get your API key here — it opens the cursor.com dashboard where you can create a key.

2. Paste it into the Enter your API key field and save with the check button.

3. Use Refresh Models in the Models list to pull the current model list from Cursor.

---

Claude Code

Claude Code doesn't use an API key. As the setting explains: "Uses your locally installed Claude Code CLI and its subscription — no API key needed."

1. Expand Claude Code and turn on the master switch. Enabling it automatically fetches the model list from the CLI.

2. Check the CLI Status card. A green check with the CLI version and its path means you're ready. Click Re-check after installing or moving the CLI.

If the CLI isn't found, the card shows install steps:

1. Install the CLI: npm install -g @anthropic-ai/claude-code

2. Authenticate once by running claude in a terminal

Options

• Custom CLI path — only needed when the CLI isn't auto-detected (e.g. a non-standard install location). Leave it empty to auto-detect.
• Permission mode — how much the agent is allowed to do:
  • Default — tools that need permission are denied (chat-only, safest)
  • Accept edits — file edits in the workspace are auto-approved
  • Plan — read-only planning mode, no changes are made
  • Bypass permissions — all tools run without asking (use with care)
• Max agent turns — how many tool-use rounds Claude Code may take per response. Between 1 and 100; default 10.

The Models list has a Refresh Models button; if the CLI can't be reached, a default model set is loaded instead.

---

Agent Workspace

When you chat with an agent-provider model, a folder picker appears in the chat input. The folder you pick is the agent's workspace — the directory it reads from and (depending on permission mode) writes to.

You can also set a canvas-level default workspace in the canvas settings, so every agent chat on that canvas starts in the same folder.

---

Use cases

• Ask questions about a codebase: point the workspace at a repo and use Plan or Default mode to explore and explain code without any risk of changes.
• Delegate real edits: switch to Accept edits and have Claude Code fix a bug or add a feature directly in your project folder, right from a chat node.
• Use your existing subscriptions: run Claude Code through the CLI's own subscription and Cursor through its API key — no extra Slashspace credits consumed for agent runs.
• Keep long tasks bounded: raise Max agent turns for complex multi-file refactors, or keep it low so quick questions never spiral into long tool loops.