How long does the integration take?

Change one base URL in an OpenAI-compatible client, or install the Python or TypeScript SDK; the quickstart routes a first request in a few minutes.

Will this lock us into another vendor?

No. The router is Apache-2.0 open source, your provider keys stay yours, and the API is OpenAI-compatible, so leaving is a change of base URL.

What happens when a provider fails?

The router falls back to the next model your policy allows and records the fallback in the trace, so an outage costs a retry, not an incident.

SDKs, API and integrations

One base URL. Every language you already write.

OpenSmartRoute is a Python library with zero dependencies, a TypeScript client typed to every endpoint, a CLI, and a hosted API that speaks the OpenAI and Anthropic dialects your clients already use. Whatever you pick, every answer carries the decision - target, confidence, alternatives, cost and a trace id.

Get a free API key Documentation

pip install "opensmartroute[yaml,server]"

npm install @opensmartroute/sdk

curl -LsSf https://opensmartroute.ai/install.sh | sh

Windows: powershell -ExecutionPolicy ByPass -c "irm https://opensmartroute.ai/install.ps1 | iex"

Four ways in

Pick the one that matches the code you have.

Python
The router as a library: targets.yaml, rules.yaml, route(), the trace. Zero runtime dependencies; optional extras for YAML, the server and OpenTelemetry.
Python SDK guide
TypeScript
Typed route, estimate and feedback calls, the /v1 surface with streaming and every platform endpoint through osr.rest. Node, Deno, Bun, edge and browsers.
TypeScript SDK guide
CLI
osr route, audit, eval, trace, events, serve and mcp from one binary. Sign in once with osr login and the commands read from the platform you chose.
Command reference
Any existing client
Point your OpenAI or Anthropic SDK at the platform's base URL. model="auto" routes; a list of models scores only those with fallbacks; stream=true works.
REST API reference

Library, sidecar or hosted API

Call it from the code you already have.

Embed the SDK in a Python service, call it from TypeScript, run the container beside your app, or use this platform. The OpenAI-compatible and Anthropic-compatible endpoints (chat, Responses, Messages, embeddings, moderations, audio, images) mean existing clients switch by changing one base URL.

Decorator SDK: Custom strategies, policy rules, signals, middleware and telemetry sinks registered with a decorator.
TypeScript SDK: @opensmartroute/sdk on npm: typed route, feedback and estimate calls, the /v1 surface with streaming, and every platform endpoint through osr.rest.
Adapters: MCP tool catalogues, A2A agent cards, LangGraph nodes and the Microsoft Agent Framework.
Enterprise builder: Auto-learning, health, budgets, tenants, guard, audit and Redis, SQL or encrypted state.
MCP server: Expose route, estimate and explain as tools to Claude, Cursor and every other MCP client.

route.pypython

pip install "opensmartroute[yaml]"

from opensmartroute import (
    Objective, RequestConstraints, RouteRequest,
    Router, load_rules, load_targets,
)
from opensmartroute.strategies import (
    default_strategies,
)

router = Router(
    load_targets("targets.yaml"),
    strategies=[
        load_rules("rules.yaml"),
        *default_strategies(),
    ],
)

req = RouteRequest(
    "Prove that sqrt(2) is irrational.",
    constraints=RequestConstraints(
        max_cost_per_1k=0.02,
    ),
    objective=Objective(quality=1.0, cost=0.4),
)
d = router.route(req, plan=True)
print(d.target.id, f"{d.confidence:.2f}")
# per-strategy scores and rationales
print(d.trace.explain())
# who was filtered out, and why
print(d.trace.policy_rejections)

Zero runtime dependencies. plan=True returns the ranked runners-up and every policy rejection with the winner.

Written for the person at the keyboard

Change one line - the base URL - and keep the client you already have.

The hosted API answers the OpenAI and Anthropic SDKs as they are. Every response carries the decision, so the trace is in your logs from the first request, not after an integration project.

Photo: Daniil Komov

Integrations

It fits the agent stack you already run.

The router imports what you have - MCP servers, A2A agents, skills, tools - as targets, and slots into the frameworks you build with as a node, a condition or an executor. Nothing has to be rewritten to be routed.

RouterPython TypeScript OpenAI SDK MCP LangGraph VS Code Docker A2A agents

Hover a spoke. Every one speaks to the same router under the same policy.

MCPImport tool catalogues, resources and prompts as targets; expose the router as an MCP server.Learn more
A2AAgent cards become agent targets with their skills and capabilities.Learn more
LangGraphA routing node and a condition that pick the next branch by decision.Learn more
Microsoft Agent FrameworkExecutors that route inside a workflow.Learn more
OpenAI function callingTool specs generated from the catalogue; parameters checked for provenance before they run.Learn more
Agent Skills (SKILL.md)Skill packages load as targets; the router selects the playbook the task deserves.Learn more

Agent frameworks

Keep the framework. Route the model.

LangChain, LangGraph, CrewAI, LlamaIndex, Semantic Kernel and AutoGen each reach the router in a few lines - most through the OpenAI-compatible endpoint with model set to auto, LangGraph and Microsoft Agent Framework with the router inside the graph. Pick yours; the snippet carries this deployment's base URL.

Through the OpenAI door

LangChain

ChatOpenAI pointed at the router; every chain call is routed and traced.

Install

pip install langchain-openai

model="auto" routes; a target id pins a model; a list of ids scores only those, with fallbacks. Every response carries the decision and a trace id.

LangChain - routedpython

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    base_url="https://osr-api.kindrock-caaab1a2.centralindia.azurecontainerapps.io/v1",
    api_key="osr_live_...",     # a workspace key
    model="auto",               # the router picks; a target id pins
)
print(llm.invoke("Summarise this clause in plain English: ...").content)

For the business

What it costs to adopt - in engineering time.

Same client, same code
One base URL
The OpenAI-compatible endpoint answers your existing SDK calls; model=auto lets the router choose.
The SDKs
One URL, every vendor
No vendor lock-in
Point your OpenAI-compatible client at one address; add, swap or retire providers in the catalogue without touching application code.
Every vendor side by side
Next eligible model answers
Failover built in
When a provider is down or over quota the request moves on; the trace says it did.
How the trace shows it

How long does the integration take?: Change one base URL in an OpenAI-compatible client, or install the Python or TypeScript SDK; the quickstart routes a first request in a few minutes.
Will this lock us into another vendor?: No. The router is Apache-2.0 open source, your provider keys stay yours, and the API is OpenAI-compatible, so leaving is a change of base URL.
What happens when a provider fails?: The router falls back to the next model your policy allows and records the fallback in the trace, so an outage costs a retry, not an incident.

Install, route, read the trace.

The hosted API needs no SDK: one curl and the decision is back. The packages are there when you want to embed the router or type the calls.

Create a free account All downloads

Free plan, no card. Fifteen thousand decisions a month with the full trace.

One base URL. Every language you already write.

pip install "opensmartroute[yaml]" from opensmartroute import ( Objective, RequestConstraints, RouteRequest, Router, load_rules, load_targets, ) from opensmartroute.strategies import ( default_strategies, ) router = Router( load_targets("targets.yaml"), strategies=[ load_rules("rules.yaml"), *default_strategies(), ], ) req = RouteRequest( "Prove that sqrt(2) is irrational.", constraints=RequestConstraints( max_cost_per_1k=0.02, ), objective=Objective(quality=1.0, cost=0.4), ) d = router.route(req, plan=True) print(d.target.id, f"{d.confidence:.2f}") # per-strategy scores and rationales print(d.trace.explain()) # who was filtered out, and why print(d.trace.policy_rejections)

Keep the framework. Route the model.

from langchain_openai import ChatOpenAI llm = ChatOpenAI( base_url="https://osr-api.kindrock-caaab1a2.centralindia.azurecontainerapps.io/v1", api_key="osr_live_...", # a workspace key model="auto", # the router picks; a target id pins ) print(llm.invoke("Summarise this clause in plain English: ...").content)

One base URL. Every language you already write.

Pick the one that matches the code you have.

Python

TypeScript

CLI

Any existing client

Call it from the code you already have.

Change one line - the base URL - and keep the client you already have.

It fits the agent stack you already run.

Keep the framework. Route the model.

LangChain

What it costs to adopt - in engineering time.

One base URL

No vendor lock-in

Failover built in

Install, route, read the trace.

One base URL. Every language you already write.

Pick the one that matches the code you have.

Python

TypeScript

CLI

Any existing client

Call it from the code you already have.

Change one line - the base URL - and keep the client you already have.

It fits the agent stack you already run.

Keep the framework. Route the model.

LangChain

What it costs to adopt - in engineering time.

One base URL

No vendor lock-in

Failover built in

Install, route, read the trace.