AI Workflow Router

Stop routing
AI work manually.

RouteWise decomposes workflows into steps, selects the best model for each, evaluates results, and generates a complete execution trace.

Get Started View on GitHub

✓ Multi-model routing✓ MCP compatible✓ Full execution trace✓ Open source

routewise run

$ routewise run "Build an AI support agent"█

·Classifyreasoning → code-gen → writing

·Plan3 steps decomposed

·Route Step 1→ Claude Sonnet 4.6

·Route Step 2→ GPT-4.1

·Route Step 3→ Gemini 3.5 Flash

·Evaluate3/3 passed

·Trace Generated$0.0087 · 4.2s

The problem with multi-model workflows

Without RouteWise

✗Claude → wrong task

✗GPT → expensive for simple work

✗Cursor → no routing visibility

✗MCP → manual tool switching

✗Scripts → fragile, no traces

With RouteWise

✓Task → Decompose → Route → Execute

✓Best model per step, automatically

✓Cost + latency tracked per step

✓Full trace: why each model was chosen

✓Works in Claude Code, Cursor, MCP

Decompose. Route. Verify.

Three phases. Every step traceable. Every decision explainable.

Plan

Decompose the task into ordered steps and classify each step type.

task graph

Route

Select the best model per step within cost, latency, and quality constraints.

model plan

Verify

Execute, evaluate output quality, and log a full trace with costs and routing decisions.

execution trace

Real routing trace

Every run produces a structured trace. See exactly which model was chosen, why, and what it cost.

·Model selected

·Reason recorded

·Cost tracked

·Latency measured

·Evaluation logged

trace.jsoncompleted

{
  "runId": "run_20260613_k7m2x1",
  "task": "Build an AI support agent",
  "status": "completed",
  "totalCostUsd": 0.0087,
  "totalLatencyMs": 4218,
  "steps": [
    {
      "stepId": "step_001",
      "type": "reasoning",
      "model": "claude-sonnet-4.6",
      "reason": "best-fit",
      "confidence": 0.92,
      "costUsd": 0.0034,
      "latencyMs": 2103,
      "evaluation": { "passed": true }
    },
    {
      "stepId": "step_002",
      "type": "code-gen",
      "model": "gpt-4.1",
      "reason": "user-preferred",
      "confidence": 0.95,
      "costUsd": 0.0021,
      "latencyMs": 1342,
      "evaluation": { "passed": true }
    },
    {
      "stepId": "step_003",
      "type": "writing",
      "model": "gemini-3.5-flash",
      "reason": "best-fit",
      "confidence": 0.88,
      "costUsd": 0.0002,
      "latencyMs": 773,
      "evaluation": { "passed": true }
    }
  ]
}

Benchmarks

Measured on a 7-step workflow: idea → plan → architecture → code → tests → docs → launch post.

78%

cost reduction vs single model

3×

faster than manual switching

94%

task completion rate

$0.009

average routed cost

Approach	Cost	Time	Completion
Single Model (Claude)	$0.042	12.4s	78%
Manual Switching	$0.028	18.7s	85%
RouteWise	$0.009	4.2s	94%

Works where you work

Native MCP server for Claude Code and Cursor. Standalone CLI for terminal workflows.

Claude Code

Cursor

MCP

OpenAI

Anthropic

Vertex AI

OpenRouterOllamaLangGraphTerminal

Open source. Apache 2.0.

RouteWise is built in public. Contribute providers, evaluators, classification heuristics, or documentation.

CLI→Router→Providers→Trace

Strict

TypeScript

125+

Tests

Providers

Models

Star on GitHub

Roadmap

Available now

✓Core routing engine

✓CLI + MCP server

✓Anthropic + OpenAI + Vertex

✓Execution traces

Coming next

Benchmark layer with auto-learn

OpenRouter + Ollama

Web UI for traces

Team routing policies