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.

Agent routing

Route between models, agents, tools and people - not only between models.

When a request needs doing rather than answering, send it to an agent; when it needs a lookup, to a tool; when it needs judgement, to a person. One catalogue, one policy, one trace, with MCP servers and A2A agents imported as targets.

Get a free API key Route this request in the playground

Free plan, no card. Open source under Apache-2.0; self-hosting is free forever.

From the benchmark

kinds of target ranked together

Models, agents, skills, personas, tools, workflows and human queues share one catalogue format, one policy engine and one trace - the marketplace publishes every kind.

The whole benchmark, with the cases the router lost

The problem

The model is one target among many. The others have no router.

A coding agent, a retrieval tool, a support persona and a human queue each get their own trigger logic, written by hand, in a different place. None of it is priced, none of it is traced, and a rule about personal data has to be written four times.

Agents and tools are chosen by hand-written conditions
No cost or quality comparison between an agent and a model
Policy and audit cover the models and nothing else

How the router answers it

Three things it does for every request.

Every kind of target in the catalogue
llm, agent, skill, persona, tool, workflow and human are first-class kinds with capabilities, a complexity band, cost and latency. The ranking treats them alike.
Import what you already run
MCP servers (tools, resources, prompts), A2A agent cards, SKILL.md packages and OpenAI function specs become targets without being rewritten.
Plans, not only picks
A request can route to a plan - persona, then skill, then model - executed step by step with the same policy, budgets and trace.

Watch agents route

Supervisor agent

Research

Analysis

Writer

Routerone policy for every agent

Large

Reasoning

Small, fast

Illustrative: each worker's calls pass through the same router and policy, and land on the model its step deserves.

A real request

Import an MCP tool server; every tool becomes a routing candidate next to the models.

Tool-parameter provenance and resource limits are enforced by the router, so an untrusted argument never reaches a tool unchecked.

Route it in the playground and read the trace

The callpython

from opensmartroute.adapters import connect_mcp_http

http, remote = connect_mcp_http(
    "https://tools.example.com/mcp",
    api_key="...", resources=True, prompts=True,
)
# every tool, resource and prompt is now a target;
# provenance checks and resource limits apply before any runs

Where this stands next to the alternatives, and how to go deeper.

Install

pip install "opensmartroute[yaml]"

The Python SDK guide

Or no install at all: the hosted API answers a plain curl with the decision, and the browser extension shows the price and the router's pick under the composer of the AI chat sites you already use.

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.

Route your first request in under a minute.

OpenSmartRoute is open source and the free plan keeps the full trace of every decision. Create a workspace, mint a key and send the request above.

Create a free account Read the benchmark

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

from opensmartroute.adapters import connect_mcp_http http, remote = connect_mcp_http( "https://tools.example.com/mcp", api_key="...", resources=True, prompts=True, ) # every tool, resource and prompt is now a target; # provenance checks and resource limits apply before any runs

Route between models, agents, tools and people - not only between models.

The model is one target among many. The others have no router.

Three things it does for every request.

Every kind of target in the catalogue

Import what you already run

Plans, not only picks

Import an MCP tool server; every tool becomes a routing candidate next to the models.

Where this stands next to the alternatives, and how to go deeper.

What it costs to adopt - in engineering time.

One base URL

No vendor lock-in

Failover built in

Route your first request in under a minute.

Route between models, agents, tools and people - not only between models.

The model is one target among many. The others have no router.

Three things it does for every request.

Every kind of target in the catalogue

Import what you already run

Plans, not only picks

Import an MCP tool server; every tool becomes a routing candidate next to the models.

Where this stands next to the alternatives, and how to go deeper.

What it costs to adopt - in engineering time.

One base URL

No vendor lock-in

Failover built in

Route your first request in under a minute.