AI features

AI features assist software work. They can be wrong. Keep tests and human review in the loop.

Cloud agent (SSE)

The coding agent runs in the cloud and streams updates over SSE to the workbench. Tool calls, reads, patches, and status appear in the agent transcript as the run progresses.

Tools the agent can use

Depending on mode and permissions, the agent may:

• Read files and list directories
• Search with grep/glob and semantic/codebase search
• Edit files and apply patches
• Use git for status, diffs, and commits when allowed
• Run shell commands under your sandbox and permission settings
• Maintain todos for multi-step work
• Read linter/diagnostics output
• Call web search
• Connect MCP servers
• Use Computer Use on supported platforms (strongest on Windows; expect caveats elsewhere)

Tab completions are local

Tab completions come from local language servers (LSP) in the workbench. They are not cloud model generations. Agent chat and cloud tools are separate from Tab.

Selection Ask and Edit

With a selection in the editor you can Ask about the highlighted code or request an Edit scoped to that selection. Prefer selection Edit for small, reviewable changes; use Agent mode for multi-file work.

Web search providers

When web search is enabled for a run, providers may include DuckDuckGo, Brave, and SerpAPI according to server configuration and your account entitlements.

Rules, skills, and MCP

Project rules, skills, hooks, and MCP servers steer and extend behavior. Dedicated pages: Tools, Subagents, Skills, MCP, and Extensions.