otto_agent

Otto is a terminal AI agent that uses what it built and judges its own work against criteria it wrote before it started. It works on a codebase, a container, a browser or a desktop. The project describes one agent loop with modes over four model vendors, a rubric-first evaluator, tiered memory, six benchmark harnesses, and 1,593 tests that need no API key. Installation is via pip install otto-cli-agent, followed by otto doctor and otto tui. INCEPTION_API_KEY is required; OpenAI, Anthropic and Gemini keys are optional and each unlocks the seats routed to that vendor. otto tui opens its setup screen on first start when nothing is configured.

The site reports measured properties. A one-tool task costs 3 model calls, with overhead flat in the length of the work. Criteria are written from the task before any attempt and shared by the loop and the judge. A write-and-run task was judged in 7 calls and 31 s, approved first time. A mutation gate is held once before an irreversible action. Claw-Eval T026 scores 0.955 with safety 1.0. Type-aware compaction keeps what the person said from being summarised, and 8/8 planted constraints were kept at 120 and 400 turns. Two-stage semantic recall over a content-addressed store achieved LoCoMo recall 96% at about 2,600 tokens per query. A greeting has no criteria, so it skips the loop; “hi otto!” costs 2 model calls. In an exercise, the agent uses what it built and the judge reads a code-written report; five of five runs walk through what they built before finishing. A golden set of 20 code, math and NP-hard tasks with real checkers scored 20/20.

Otto is MIT licensed and built by Siddharth Priyadarshi. otto_agent is maintained by siddharth23p.