Compare · MLflow AI Gateway and Selan

Is there an MLflow AI Gateway alternative for a team's Claude Code?

Yes, if the traffic is your team's coding tools and nobody wants to run a server. MLflow AI Gateway is built into the MLflow Tracking Server, which you run yourself. Claude Code reaches it through ANTHROPIC_BASE_URL and keeps each developer's own Anthropic sign-in. Every request becomes an MLflow trace. Selan is a hosted gateway that each developer signs in to once. If your team already traces its AI apps in MLflow, MLflow AI Gateway is the better answer.

How do MLflow AI Gateway and Selan compare?

	MLflow AI Gateway	Selan
What it is	A gateway inside the MLflow Tracking Server, with traces, usage charts, budgets, guardrails, traffic splitting and fallbacks	A hosted gateway, with a CLI that launches Claude Code, Codex and OpenCode through it
Where it runs	Your own server, started with `mlflow server` on a SQL database; MLflow names managed versions on Databricks and AWS	A hosted service, built and hosted in the EU. Nothing to run
Coding tools	Setup pages for Claude Code, OpenAI Codex, Gemini CLI and Hermes Agent	Claude Code, Codex, OpenCode and Claude Desktop
Claude for Claude Code	Each developer's own Anthropic sign-in, passed through the gateway	The Claude and ChatGPT accounts and API keys your company connects
A developer's credential	Their own Anthropic login; with MLflow's authentication on, also an MLflow user allowed to use the endpoint	One sign-in with Google or Microsoft through `selan login`
Usage	A dashboard of requests, latency, errors, tokens and cost by model or provider, for all endpoints or one	One usage record per person, across every tool
Limits	Dollar budgets per day, week or month, global or per workspace, that send an alert or reject with a 429	A spend limit per person, by day, week or month in UTC, on pay-per-use keys
Content checks	Guardrails: an LLM judge that blocks or redacts requests and responses	Secret scanning replaces a secret it recognises before the request leaves
Cost	Free and open source under Apache 2.0, on infrastructure you run	Free for up to three seats, no card

MLflow AI Gateway records what a team's coding agents send, on a server you run. Selan is a gateway a team signs in to for its coding tools.

When is MLflow AI Gateway the better answer?

When your team already uses MLflow for its agents and LLM apps. Requests through the gateway become MLflow traces, and MLflow's evaluation runs on those same traces.

Each trace holds the request, the response, the token counts and the latency. That shows what a long agent session actually sent and where its tokens went.

It is free under the Apache 2.0 licence, backed by the Linux Foundation, and runs where you put it. The gateway needs no separate service: it ships in the tracking server.

Its guardrails are instructions you write in plain language for an LLM judge. A guardrail runs before the model or after it, and either blocks the call with a 400 or redacts what it flags.

It also has setup pages for Gemini CLI and Hermes Agent. For your own apps it reaches OpenAI, Anthropic, Amazon Bedrock, Google Gemini, Azure OpenAI, Mistral, DeepSeek and more, with traffic splitting and ordered fallbacks.

If no team runs coding tools through a gateway, you do not need Selan.

Whose Anthropic account does Claude Code use through MLflow?

The developer's own. MLflow's Claude Code page says each developer authenticates with their own Anthropic subscription. Claude Code keeps the login stored in ~/.claude, and the gateway passes each request on to Anthropic.

MLflow calls this visibility without central key management. The gateway adds traces, budgets and guardrails on top, and the developer's Anthropic account stays theirs. The Claude Code route arrived in MLflow 3.12.0.

Codex differs. MLflow's Codex page says to sign in with an OpenAI API key instead of a ChatGPT subscription. With MLflow's authentication on, Codex also sends an MLflow user name and password in an X-MLflow-Authorization header, which MLflow never forwards to OpenAI.

Good to know: a trace keeps the full request and response, so what a session sends is stored on your MLflow server. Guardrails that check the model's answer do not run on streaming requests; MLflow evaluates only non-streaming responses.

How do MLflow AI Gateway budgets work?

A budget policy is a dollar amount over a day, a week or a month. Days reset at midnight UTC, weeks on Sunday and months on the 1st.

An alert policy fires a webhook once per window, and requests carry on. A reject policy answers later requests with an HTTP 429. The request that crosses the line still completes.

A policy covers the whole gateway, or one workspace when workspaces are on. With authentication on, only admins can create, change or delete one.

Good to know: by default each server process tracks spend on its own, so several workers together can pass the limit. A shared limit needs Redis, set in MLFLOW_GATEWAY_BUDGET_REDIS_URL. Token and cost charts can be empty for some providers and models, depending on LiteLLM's pricing data.

When is Selan the better answer?

When the traffic is a team's coding tools and nobody wants to run a server. A developer runs selan login once, then selan claude, selan codex or selan opencode. Claude Desktop goes through the same gateway with one settings file.

Claude Code, Codex, OpenCode and Claude Desktop go into one usage record per person. A limit per person applies across every pay-per-use key.

An owner connects the company's accounts and keys once, and switches on the models the team may use. No provider key reaches a laptop. With secret scanning on, a secret it recognises is replaced before the request leaves.

The trade-off is that requests travel through Selan, so if Selan is unavailable those requests fail. Your accounts can still be used directly.

Which providers serve which models is in models and providers, limits per person are on the spend page, and scanning is in the secret scanning docs.

How do you move Claude Code from MLflow AI Gateway to Selan?

Remove the ANTHROPIC_BASE_URL that points at your MLflow server, from your shell profile or from the env block of ~/.claude/settings.json. For Codex, remove openai_base_url and env_http_headers from ~/.codex/config.toml.

Then run selan login and selan claude. Connect the company's Anthropic accounts or keys in Selan and switch on their models. You can keep MLflow for your own apps' traces and run the team's coding tools through Selan. Keeping provider keys off laptops is covered in Claude Code without API keys on laptops, and the kinds of gateway in Claude Code gateways.

Start for free

Free for up to three seats, and the Free plan needs no card.

The short version

MLflow AI Gateway is built into the MLflow Tracking Server you run, free under Apache 2.0.
Claude Code reaches it through ANTHROPIC_BASE_URL, keeps each developer's own Anthropic sign-in, and every request becomes a trace.
Budgets are in dollars, for the whole gateway or one workspace, and alert or reject with a 429.
Selan is hosted in the EU: one sign-in per person, no provider key on the laptop, and a spend limit per person.

What MLflow AI Gateway does is as MLflow's documentation (the AI Gateway overview and quickstart, the Claude Code and OpenAI Codex pages, usage tracking, budget alerts and limits, guardrails, traffic routing and authentication), its AI Gateway page, its 25 May 2026 post on routing Claude Code through the gateway and the mlflow/mlflow repository (release 3.18.0) described it, read on 6 November 2026. It changes: check the date before relying on it. What Selan does is as its CLI and docs stated it on that date.