Skip to content
LoopCodeDocs

Use LoopCode

How LoopCode works

The agent loop behind every LoopCode conversation — tools, permissions, evidence and proof — and where the work happens.

The agent loop

Each message you send starts a turn. In a turn, the model works in a loop:

  1. Gather context — it reads files, searches the code, asks the language server for definitions and references, looks at git state, and loads the project's instructions and memory.
  2. Act — it edits files, runs commands in a terminal, drives a browser, or delegates part of the job to a subagent.
  3. Check — it runs the tests, the typecheck, the lint and the app itself, and reads the results.
  4. Repeat until the work is done and proven, then report.

You see every step as it happens, and you can interrupt, steer or queue a follow-up at any time.

Tools

The model can only affect your project through tools. LoopCode's built-in tools read and search files, edit them, run shell commands and long-lived terminals, read git state and make commits, fetch web pages, drive a browser, build Word, PowerPoint and Excel files, and talk to language servers. MCP servers add more. The full list is in the tools reference.

Permissions

Before a tool call that changes something runs, LoopCode decides whether it may go ahead. That depends on the permission mode — by default Auto-review, where routine work proceeds and a reviewer model stops risky actions to ask you — and on the rules in your settings. File edits are always shown as diffs you can undo.

Proof of done

Everything the agent checks during a turn is kept as evidence: a test run, a typecheck, a build, a screenshot, a command's output — marked passed or failed from the tool's own result, not from what the model says. Evidence goes stale as soon as the code changes again.

Before a turn that changed code can end, the agent states what you asked for as checkable criteria and cites fresh, passing evidence for each. LoopCode accepts that only when the project's own checks passed after the last edit. See Proof of done.

Context

The model works within a context window. LoopCode keeps it efficient: old tool output is cleared once it has been used, long conversations are summarised automatically when the window fills, and subagents do heavy reading in their own windows and return only a summary. See Context and compaction.

Where it runs

The LoopCode engine runs on your computer — inside the desktop app, the CLI, or the VS Code extension — and works directly on your files with your tools. Your conversations are stored locally. The model itself is hosted: each request sends the relevant prompt and code context to LoopCode's model service. See Data and privacy.

Memory across conversations

LoopCode starts every conversation with your project's instruction files (LOOPCODE.md, AGENTS.md, CLAUDE.md), what you have asked it to remember, and notes it learned from earlier sessions. See Instructions and memory.