What is myles?

Myles is a local proxy designed to protect developers from runaway AI agent costs. When autonomous agents get stuck in loops, retrying failed tool calls endlessly or running impossible lookups across many parallel sessions, the bills can reach thousands of dollars in hours. Myles sits between your agent and the AI provider, watching requests in real time and blocking the ones that would waste money. The project installs with a single pip command and requires no changes to your existing agent code. You start it with a demo command that needs no API key, then point your agent at a local address. By default Myles runs in observe mode, logging what it sees and reporting what it would have blocked, without actually interfering. Once you trust the detection, you enable enforcement with a budget cap and it starts refusing costly runaway calls before they reach the provider, meaning blocked requests cost zero dollars. Myles catches three specific problems. It detects runaway loops using semantic novelty detection, flagging calls that stop carrying new information even when the agent rephrases the same failing request or spreads it across tools. It meters real dollar costs from actual token usage and enforces a hard cap. It also catches fan-out, where the same doomed intent appears across many concurrent sessions, containing it fleet-wide rather than per process. The detector measures information novelty per step and uses IDF weighting to learn your boilerplate so common operations are ignored while the real intent is evaluated. The project is at version 0.1, an early MVP seeking feedback. Enforcement has been validated end to end. Honest limitations are noted: cross-run detection uses a coarse intent key that can be sensitive on short intents, and streaming is buffered rather than true pass-through. The project uses only the Python standard library with zero dependencies, requires Python 3.9 or later, and is released under Apache-2.0.