langchain-router

Your agent doesn't need the expensive model for every call.

Most calls are just the model picking which file to read next or which pattern to search for. A smaller model does that fine. This middleware detects when the agent is doing that kind of work and routes to a fast model automatically.

Quick Install

pip install langchain-router

🤔 What is this?

Agent sessions have a pattern. The user says something, the agent thinks about it (planning). Then it reads files, searches code, runs commands (execution). Sometimes something breaks (recovery). Then the user says something again.

Planning and recovery need the primary model. Execution doesn't. RouterMiddleware detects which phase the agent is in and routes accordingly.

No ML. No embeddings. No classifier. Just message types.

from langchain.agents import create_agent
from langchain_router import RouterMiddleware

agent = create_agent(
    model="anthropic:claude-sonnet-4-6",
    tools=[...],
    middleware=[RouterMiddleware(fast="anthropic:claude-haiku-4-5-20251001")],
)

With CollapseMiddleware

from langchain_collapse import CollapseMiddleware

middleware = [
    CollapseMiddleware(),
    RouterMiddleware(fast="anthropic:claude-haiku-4-5-20251001"),
]

flowchart TB
    A["📥 37 messages"] --> B["CollapseMiddleware"]
    B --> C["📥 9 messages"]
    C --> D["RouterMiddleware"]
    D --> E{"phase?"}
    E --> |"execution  ·  83%"| F["⚡ GLM-5"]
    E --> |"planning"| G["🧠 Opus 4.6"]
    E --> |"recovery"| G

    style A fill:#ff6b6b,stroke:#e03131,color:#fff
    style B fill:#339af0,stroke:#1c7ed6,color:#fff
    style C fill:#339af0,stroke:#1c7ed6,color:#fff
    style D fill:#51cf66,stroke:#2f9e44,color:#fff
    style E fill:#fff3bf,stroke:#f59f00,color:#333
    style F fill:#20c997,stroke:#099268,color:#fff
    style G fill:#845ef7,stroke:#7048e8,color:#fff

Verified with real API calls

81 real API calls across 9 model pairs, 4 providers. Each pair ran a 9-call agent session: planning, 6 execution turns, recovery from a failed tool call, and re-planning. Every call routed correctly. Full per-call details.

Patterns tested: same-provider (Anthropic→Anthropic, OpenAI→OpenAI), cross-provider (Anthropic→Z.ai, OpenAI→MiniMax), frontier→open (Opus→GLM-5, GPT-5.4→MiniMax), open→open (GLM-5→MiniMax).

Cost impact

In a typical coding session (8 file reads, 4 greps, 1 edit, 1 test failure, 1 fix, 1 re-test, 2 user messages), 83% of model calls are execution phase and route to the fast model.

Token counts assumed for cost projection (8k input, 1.5k output per call). Pricing as of April 2026. See benchmark for methodology.

On false positives

Recovery detection checks ToolMessage.status first (set by the agent when a tool call fails), then falls back to content scanning for error, traceback, exception, failed. Code containing those words (like def handle_error) routes to the primary model. That's the safe direction: more capability than needed, never less.

📖 Documentation

• Source (single file, ~170 lines)
• Benchmark (cost projections across providers)
• Verified Results (81 real API calls, 9 model pairs)
• Tests (unit, integration, and property based invariant tests)

💁 Contributing

git clone https://github.com/johanity/langchain-router.git
cd langchain-router
pip install -e ".[test]"
pytest

📕 License

MIT