Rebel Router
An open-source local proxy tool (MIT license) developed by the larebelion team, operating between coding agents and model providers. It uses intelligent routing to automatically select the optimal model for each request and supports cost tracking and budget control.
Timeline (last 90 days)
Rebel Router is an open-source project built by the larebelion team under the MIT license, operating as a locally running proxy that selects the most suitable model for each request.
Rebel Router runs as a local proxy (127.0.0.1:8787) between the coding agent and model providers, supporting requests from Claude Code (Anthropic Messages format) and OpenCode (OpenAI Chat format).
The providers supported by Rebel Router include Anthropic, OpenAI, xAI (Grok), Google Gemini, DeepSeek, and local Ollama.
Rebel Router consists of three components: Broker (model catalog and ranking), Router (classifier and proxy core), and FinOps (budget and cost accounting).
Rebel Router categorizes requests into six classes—orchestration, complex_code, bulk_code, unit_tests, qa_automation, and documentation—and assigns a complexity score from 0 to 1.
Rebel Router can invoke Venice's JEV decision model (Beta stage) for classification, falling back to a local rule engine if an error occurs, the response is off-topic, or the response exceeds 800ms.
Rebel Router's budget control mechanism: at 80% of budget it favors cheaper models; at 100% it downgrades to the cheapest acceptable model or returns HTTP 402.
In Broker ranking, input and output tokens are priced with a 3:1 blended ratio, and the top three candidates are selected by a weighted combination of quality, price, and speed.
The benchmark consists of 12 real tasks (two per category), all scored deterministically, completed on October 11, 2026, covering 10 models, with one attempt per task, at a total cost of $1.78.
The intelligent routing solution costs $0.17, while using Opus 5.5 throughout costs $0.96, making intelligent routing 82% cheaper.
4 more timeline events
All Facts (14)
Rebel Router is an open-source project built by the larebelion team under the MIT license, operating as a locally running proxy that selects the most suitable model for each request.
50%UnverifiedRebel Router runs as a local proxy (127.0.0.1:8787) between the coding agent and model providers, supporting requests from Claude Code (Anthropic Messages format) and OpenCode (OpenAI Chat format).
50%UnverifiedThe providers supported by Rebel Router include Anthropic, OpenAI, xAI (Grok), Google Gemini, DeepSeek, and local Ollama.
50%UnverifiedRebel Router consists of three components: Broker (model catalog and ranking), Router (classifier and proxy core), and FinOps (budget and cost accounting).
50%UnverifiedRebel Router categorizes requests into six classes—orchestration, complex_code, bulk_code, unit_tests, qa_automation, and documentation—and assigns a complexity score from 0 to 1.
50%UnverifiedRebel Router can invoke Venice's JEV decision model (Beta stage) for classification, falling back to a local rule engine if an error occurs, the response is off-topic, or the response exceeds 800ms.
50%UnverifiedRebel Router's budget control mechanism: at 80% of budget it favors cheaper models; at 100% it downgrades to the cheapest acceptable model or returns HTTP 402.
50%UnverifiedIn Broker ranking, input and output tokens are priced with a 3:1 blended ratio, and the top three candidates are selected by a weighted combination of quality, price, and speed.
50%UnverifiedThe benchmark consists of 12 real tasks (two per category), all scored deterministically, completed on October 11, 2026, covering 10 models, with one attempt per task, at a total cost of $1.78.
50%UnverifiedThe intelligent routing solution costs $0.17, while using Opus 5.5 throughout costs $0.96, making intelligent routing 82% cheaper.
50%UnverifiedThe intelligent routing solution passes 9/12 (75%) tasks and Opus passes 11/12 (92%); of the three failures, two unit test tasks and one QA task were routed to Claude Haiku 5.5, whose responses exceeded the 8,000-token limit.
50%UnverifiedIntelligent routing reduces average time per task from 31.7 seconds to 19.5 seconds.
50%UnverifiedThe purely rule-based classifier without JEV correctly categorized 9 out of 12 tasks.
50%UnverifiedNo Claude Code hook can change the session model; the interface always displays the model requested by Claude Code, while the router status bar shows the model that actually responded.
50%