The Agent

Switch the composer to Agent mode, describe a task in plain English (e.g. “add a /health route and a test for it”), and press Send. The agent works through the task step by step, showing you its reasoning, what it’s doing, and the result of each step as it goes — you’re never waiting on a black box.

What it can do

• Read files (whole file, or a specific line range), list directories, search text across the workspace (grep), find files by pattern (glob), and view your current git diff.
• Create new files, edit existing ones (single find/replace, or several find/replace edits to one file batched into a single step), and run shell commands.
• Call any configured MCP server tools too.

Safety and control

• Approval gates + end-of-run batch review. Every file write and terminal command asks for approval before the agent proceeds — but file edits no longer land on disk one by one. They’re held in memory and presented at the end of the run as one batch review: a card of per-file diffs, each with its own Accept/Reject, plus Accept all / Reject all. Rejected files never touch disk at all. The one exception: if the agent runs a terminal command mid-run, edits made before that point are written first (the command needs real files to act on) — those were each individually approved already.
• One-click revert. Every agent run is checkpointed. Revert undoes everything it did — restores edited files to how they were, deletes any files it created — in one step.
• Bounds. A max-steps setting caps how many steps a run can take before stopping on its own (default: unlimited — it stops only when the model says it’s done, or you press Stop yourself).
• Works with any model. No reliance on a model’s built-in “function-calling” feature — every model speaks one universal text-based tool protocol, so a small local Ollama model works the same way as a large hosted one. Agent quality still depends on the model, though: a 7B local model won’t reason as well as a top hosted one.

MCP servers

MCP (Model Context Protocol) servers give the agent extra tools beyond the built-in file/terminal ones. Configure them by hand-editing the shared ~/.xpreiide/config.yaml:

mcpServers:
  filesystem:
    command: npx
    args:
      - -y
      - "@modelcontextprotocol/server-filesystem"
      - /path/to/allowed/directory

Once configured, a server’s tools are automatically available in Agent mode, labeled MCP: <tool> (<server>) in the approval card and step log. Every MCP tool call requires your approval, exactly like a file write or terminal command. If a configured server fails to start, it’s silently skipped — its tools just won’t show up; everything else keeps working.