Warp Lets Developers Plug Any AI Model Into Their Terminal Agent
Warp's AI agent now accepts any OpenAI-compatible endpoint, letting you swap in OpenRouter, LiteLLM, or a private gateway and switch models with a slash command.
Takeaways
- Warp Agent now accepts any OpenAI-compatible custom inference endpoint — OpenRouter, LiteLLM, internal gateways, or tunneled local models.
- New model aliases let you name endpoints and switch between them mid-session using
/model [alias]. - Free for individuals and teams up to 10 people; no Warp AI credits consumed when using your own endpoint.
- Requests still route through Warp's backend (for the agent harness), but API keys are never stored server-side.
- Custom endpoints only work for local agents — cloud agent runs still use Warp's bundled inference and credits.
- A fully client-side Rust harness for local models (no server routing) is on the roadmap, along with Agent Client Protocol support.
Warp just shipped one of its most-requested features: the ability to plug any OpenAI-compatible inference endpoint directly into the Warp Agent. Set an endpoint URL, give it a memorable alias, and switch between models mid-session with /model [alias]. No more being locked to whatever models Warp bundles by default.
The problem this actually solves
Warp's agent is deeply integrated with your terminal state: it reads your indexed codebase, environment variables, MCP servers, rules, and notebooks without any extra wiring. That's the part developers want to keep. What they didn't want was having no say in which model sits underneath it all.
Custom endpoints are for developers who want control over inference, including the ability to experiment with providers and models Warp does not natively support yet. Bundled inference remains available for developers who want Warp to handle model access and infrastructure for them.
What you can actually plug in
You can now power Warp's agent experience with your own OpenAI, Anthropic, or Google API key, or connect Warp to an OpenAI-compatible endpoint such as OpenRouter, LiteLLM, z.ai, an internal gateway, or your own hosted inference setup.
Anything that speaks the OpenAI Chat Completions API (POST /v1/chat/completions) is fair game.
- OpenRouter , connects your OpenRouter URL directly to the Warp Agent, providing immediate access to models from Anthropic, OpenAI, and DeepSeek without leaving the command line.
- LiteLLM , a self-hosted proxy that exposes a unified OpenAI-compatible API across dozens of providers.
- Internal gateways , corporate AI gateways with logging, redaction, or access control, as long as they're reachable at a public URL.
- Local models via tunnel , Ollama or vLLM running on your machine, exposed through ngrok or Cloudflare Tunnel.
How the request actually flows
It's worth understanding the plumbing here, especially if you're security-conscious. Your local Warp client pulls your endpoint URL and API key from your device's secure storage and sends them up to Warp's backend. Warp's agent harness assembles the full request (system instructions, conversation context, tools) and uses your key in-flight to call your configured endpoint. Your endpoint's response then streams back through Warp's backend to your client.
Your API key passes through Warp's servers each time you send a request, but Warp never stores it there , it's used only in-flight to call your endpoint, then discarded. The agent harness itself still runs server-side, which is why the request routes through Warp at all. A custom endpoint swaps the upstream destination; it doesn't change where the orchestration logic runs.
One hard constraint worth knowing:
your endpoint must be reachable at a public URL. Requests route through Warp's servers, so Warp must be able to reach your endpoint over the public internet. localhost, private or internal network addresses, and internal-only services are rejected.
Model aliases and the /model command
The new model alias feature is the quality-of-life piece that makes this practical day-to-day. Instead of pasting a full endpoint URL every time you want to switch, you define a short alias once and call it with /model [alias] from anywhere in the agent. This is especially useful if you're routing through OpenRouter and want to flip between, say, a fast cheap model for scaffolding and a frontier model for the hard parts.
Setting it up
- Open Warp and go to Settings → AI.
- Search for
inference endpointto jump to the configuration section. - Add your endpoint base URL (e.g.
https://openrouter.ai/api/v1) and your API key. - Specify the model identifier(s) you want to route through it.
- Set a short alias you can remember.
- Switch to it anytime with
/model [your-alias].
For local models, you'll need to tunnel them first. With Ollama running on port 11434, for example:
ngrok http 11434
# Then use the resulting https://*.ngrok-free.app/v1 as your base URLWhat it costs (and what it doesn't)
BYOK and arbitrary endpoint support are available for individuals and companies of up to 10 employees. It's login-gated, but free to use within that scope. When you use your own endpoint, Warp does not consume your AI credits. Instead, all inference costs are billed directly through your endpoint provider account.
For companies with more than 10 people, BYOK and custom endpoint usage needs to go through Warp's Business or Enterprise plans. Since the customer is providing the model provider, API key, or endpoint, BYOK and custom endpoint usage will consume Warp credits at a significantly lower rate than usage where Warp also provides the underlying model inference.
One gotcha: Warp's Auto routing mode always uses Warp credits regardless of your custom endpoint configuration. You need to explicitly select your endpoint-routed model from the picker to avoid drawing from your Warp credit balance.
The limits right now
- Custom AI model selection is not available for cloud agent runs; custom endpoints remain available for local agents and non-cloud conversations.
- Each user adds their own endpoint locally; centrally configured, admin-managed endpoints for teams are not yet available.
- ZDR (Zero Data Retention) guarantees that Warp holds with its contracted providers do not extend to custom endpoints , data handling is governed by your chosen provider's own policies.
What's coming next
Warp is planning to add a lightweight Rust client harness to their open-source repo, so Warp can connect directly to local models without routing through Warp's servers. They also plan to support Agent Client Protocol, so developers can bring other agent harnesses into Warp's terminal UI. These paths will be fully client-side and will not require login.
The bigger picture here is Warp positioning itself as the interface layer for whatever model stack you're running, rather than a closed system tied to its own inference. Longer term, developers and teams will not use one model, one provider, or one harness. They will use different setups depending on the task, cost, latency, security requirements, and execution environment. Warp's goal is to be the best interface for that world. The official docs have the full configuration reference if you're ready to wire something up.