Agents

Weaver does not ship its own model. It drives coding agents you already use. The daemon runs each agent on your computer, translates its output into one transcript format, and sends that to every client.

Which agents are on

Agent	In the app	What runs
Claude Code	On	The Claude Code binary bundled with Weaver, through Anthropic’s Agent SDK.
Codex	On	The Codex CLI you install. Weaver looks on your PATH and at ~/.local/bin/codex.
Pi	On	Pi runs as a library inside the Weaver daemon. Nothing to install.
Cursor	On	The Cursor SDK bundled with Weaver, in local mode.
OpenCode	Built, off	Off in production builds.
Copilot	Built, off	Off in production builds.

The pickers show Claude Code as “Claude”.

OpenCode and Copilot adapters exist in the code but production builds hide them. A hidden agent does not start its runtime, read its credentials, or import its history. The Previews pane that can flip these flags locally is not available in production builds.

Signing in

Weaver reads each agent’s own login. It does not keep a separate account for Claude, OpenAI or Cursor. Open Settings → Agents, pick the agent’s tab, and look at Connections. Each row shows whether the runtime is installed and whether you are signed in. If something is missing, the row offers the fix. Sign-in steps run in a terminal inside that row.

Claude Code

• Weaver uses your Claude Code login (the ~/.claude config, or CLAUDE_CONFIG_DIR if you set it).
• It also accepts ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN or CLAUDE_CODE_OAUTH_TOKEN from the environment or from project settings.
• If you are not signed in, pick one of three sign-in actions: Use Claude subscription (Pro, Max, Team, Enterprise), Use Anthropic Console (API billing), or Use organization SSO. Each runs claude auth login with the matching flag.
• The Connections row shows the method, source, email, organization and plan that Claude Code reports.

Codex

• Install the Codex CLI first. If Weaver cannot find it, the row offers Install Codex CLI, which installs the official CLI into ~/.local/bin on the computer running Weaver.
• Weaver reuses the credentials Codex already stores in ~/.codex. If you have used Codex before, there is nothing else to do.
• If you are not signed in, pick Log in with ChatGPT (runs codex login --device-auth) or Use API key. The API key option opens a hidden prompt on the computer running Weaver.

Pi

• Pi supports many model providers. Weaver shows each provider Pi can use and whether it is ready.
• Use Choose provider in Connections to sign in to a provider. Some providers use OAuth only.
• Credentials stay in Pi’s own store under ~/.pi.

Cursor

• Cursor is different. Its SDK does not reuse the Cursor app login. It needs a Cursor dashboard API key.
• Pick Configure API key in Connections. A hidden prompt opens on the Mac running Weaver. Paste the key there. Weaver saves it to a file only your user can read, inside the Weaver data directory, and never sends it over the wire.
• If CURSOR_API_KEY is set in the daemon’s environment, Cursor uses it. A saved key wins over the environment variable.
• The key prompt needs a macOS host. On a Linux host, set CURSOR_API_KEY for the daemon instead.

You can also run /login and /logout from the composer. See Slash commands and references.

Approvals are bypassed

Weaver has no approval prompts today. Every agent runs its tools, edits files and runs shell commands without asking you first.

• Claude Code runs with permissionMode: "bypassPermissions".
• Codex runs with approval policy never and sandbox danger-full-access.
• Pi has no permission system.
• Cursor runs in the SDK’s local mode. Weaver shows no approval step for it.

Run agents only in projects where you accept that. Use a worktree when you want the changes kept apart from your main checkout. See Worktrees and projects.

One exception: for Claude Code and Pi, Weaver blocks shell commands that would stop, restart or kill the Weaver daemon that hosts the session.

Claude Code and Pi can still ask you a question in the middle of a turn. That is a structured question with choices, not a permission request.

What each agent supports

Every control in Weaver is gated by a flag the agent’s adapter declares. If an agent cannot do something, the button does not appear. The table shows the flags as the adapters declare them today.

Feature	Claude Code	Codex	Pi	Cursor
Model picker	Yes	Yes	Yes	Yes
Thinking or effort level	Yes	Yes, per model	Yes, per model	Yes, per model
Separate thinking on/off	No	No	No	Yes, per model
Context window choice	No	No	No	Yes, per model
Speed	Yes, on models with fast mode	Yes, per model	No	Yes, per model
Legacy models in Settings	Yes	No	No	No
Image attachments	Yes	Yes	Yes	Yes
File and video attachments	Yes, as a saved local path	Yes, as a saved local path	Yes, as a saved local path	Yes, as a saved local path
Agent slash commands	Yes	Yes	Yes	Yes
MCP servers	Yes	Yes	No	Yes
MCP management (/mcp)	Yes	Yes	No	No
Questions during a turn	Yes	No	Yes	No
Steer a running turn	Yes	Yes	Yes	No
Queue a follow-up	Yes	Yes	Yes	Yes
Cancel a turn	Yes	Yes	Yes	Yes
Force stop a stuck turn	Yes	Yes	Yes	No
Fork that keeps the agent’s own context	Yes	Yes	Yes	No
Resume another agent’s conversation	Yes	Yes	Yes	Yes
Import sessions you ran before Weaver	Yes	Yes	Yes	No
Background tasks	Shell, subagent, workflow, monitor, task	Shell, subagent	No	Subagent
Plan and todo display	Yes	Yes	No	Yes
Launch environment variables	Yes	Yes	No	No
Token counts	Yes	Yes	Yes	Yes
Context left	Yes	Yes	Yes	No
Account quota in Settings	Yes, experimental	Yes	No	No
Redeem a rate-limit reset credit	No	Yes	No	No
Inline visualizations	No	Yes	No	No

Notes:

• Attachments. Images go straight to the agent. Other files and videos are saved on the host and the agent gets the path, then reads them with its own tools. SVG images also go by path.
• Force stop. Cursor runs the turn on Cursor’s side. If cancel does not land, Cursor may keep working and Weaver cannot stop it. You still get the force-stop button so the session can leave the running state.
• Fork. Cursor has no native fork. Weaver can still fork a Cursor session by copying the conversation into a new one. See Sessions.
• History import. Cursor import is a Weaver gap, not a Cursor limit. It is tracked for a fix.
• Context left. Cursor reports tokens but not the size of its context window, so Weaver cannot show a percentage.
• Pi and MCP. Pi does not support MCP by design.
• Launch environment. Set variables per agent in Settings → Agents → Environment. Agents that cannot take them say why. Pi runs inside the daemon, so per-agent variables would leak into other agents.

Pick a default agent

Set the agent that new tasks suggest in Settings → General → New tasks. Override it per project in Settings → Projects & New Tasks. Every new task still asks you to confirm its agent and project. No suggestion leaves the choice open.