OmniRoute

OmniRoute is a self-hosted AI gateway that puts a large catalog of model providers behind one OpenAI-compatible endpoint on localhost. Point any tool that speaks the OpenAI, Anthropic, Gemini, or Responses API at that endpoint and calls are routed on your behalf; a fresh install answers immediately with keyless free providers pre-wired, before any credentials are configured.

Routing is organised around combos: a chain of models the gateway moves across when quota runs out, a provider fails, or costs rise. Nineteen strategies are available per combo step — among them priority, fill-first, weighted, round-robin, least-used, cost-optimized, headroom, cache-optimized, and last-known-good-provider — and a set of auto model IDs builds a virtual combo from your connected providers scored for balance, coding quality, latency, cost, or quota headroom. Resilience is layered: a provider circuit breaker, a per-key connection cooldown that honours Retry-After, and a per-model lockout that never disables a whole connection.

A compression pipeline sits in front of the provider and needs no client changes. Twelve engines run in order and can be toggled per combo — session deduplication, context archiving behind retrieve markers, whitespace trimming, tool-result filtering, tabular JSON compaction, extractive relevance scoring, rule-based prose compression, summarization with progressive aging, machine-learning semantic pruning, heuristic token pruning, and an experimental context-as-image encoding. Code blocks, URLs, and structured data are preserved. The default stacked combination is documented as follows:

Preset	Documented savings	Intended for
Lite	~15%	Always-on safe default
Standard	~30%	Daily coding
Aggressive	~50%	Long tool-heavy sessions
Ultra	~75%	Maximum savings
Tool-result filtering	60–90%	Shell, test, build, and git output
Stacked	78–95%	Mixed prompts and tool logs

The gateway ships a CLI that can launch a supported coding tool through OmniRoute in one command, injecting credentials per process without writing config files, or write that tool’s own configuration interactively. A remote mode runs the CLI locally against an OmniRoute instance on a server. Beyond routing, it exposes an MCP server and an A2A agent endpoint, so an agent can drive OmniRoute itself, and runs as a server, a desktop application, a progressive web app, or under Termux.

Keys are stored locally with AES-256-GCM encryption, and a dashboard reports usage, quota, savings, and p95 latency, including a page that tracks remaining free-tier budget across the catalog.

Features

• One endpoint: an OpenAI-compatible local URL in front of a large registered provider catalog
• Zero-config start: keyless free providers answer an auto request on a fresh install
• Combos: chains of models with automatic movement to the next healthy target on quota, failure, or cost
• 19 routing strategies: priority, fill-first, weighted, round-robin, least-used, cost-optimized, headroom, cache-optimized, and more
• Three resilience layers: provider circuit breaker, per-key cooldown honouring Retry-After, and per-model lockout
• 12-engine compression: context reduction that requires no client changes, with code, URLs, and structured data preserved
• CLI launcher: run or configure a supported coding tool through the gateway, with a dry-run preview
• Remote mode: drive a gateway hosted elsewhere from the local CLI
• MCP and A2A: an agent can call the gateway’s own toolset and other agents over A2A
• API compatibility: OpenAI, Anthropic, Gemini, and Responses API shapes at the same /v1 path
• Local-first storage: keys held locally with AES-256-GCM encryption
• Dashboards: usage, quota, savings, p95 latency, and remaining free-tier budget