Documentation¶

millivolt is one self-hosted service for routing OpenAI-compatible traffic, seeing what happened, and controlling the operator surface. Start with the path that matches what you need to do.

Start here¶

• Run it locally - Docker Compose first, source second.
• Connect a client - base URL, routing headers, and a working curl.
• Tour the dashboard - the views, charts, request drawer, and Settings.
• Read the protocol - endpoints, upstream routing, retries, model discovery, and conversations.

Guides¶

• Upstream compatibility and adapters - OpenAI-compatible, Anthropic Messages, Cursor Connect, and token refresh.
• Reverse proxy - protected ingress, credentials, and NGINX.
• MCP server - the 22-tool operator API for LLM-driven analysis.
• Container operation - image selection, volumes, backups, and shutdown.

Reference¶

• Operations - configuration, persistence, routes, storage, performance, and limits.
• Performance and resource usage - current measured tables, method, commands, and limitations.
• Architecture - ownership, invariants, and where each behavior lives.
• Versions and images - version identity, promotion, tags, and verification.

Project¶

• What you get - the short feature overview.
• Contributing - development, checks, branch promotion, screenshots, and release policy.
• Security - exposure boundary, reporting, and supported versions.
• Releases - published versions and notes.
• License - MIT terms and dependency notices.

Verification and limitations¶

The generated configuration example documents settings. internal/config owns defaults, validation and Settings metadata; local proxy.yaml is not a public reference or a test fixture.

Current commands live in Contributing. Runtime limits live in operations; measurement caveats live in performance. The architecture map connects invariants to their shared implementation and regression coverage. Past development notes are not an operating guide or a capacity guarantee.