Le Répertoire des AgentsFR · EN · ES · JA · HI

“My agents spend out of control and the API bill explodes”

Developers and small teams that have already received a hefty API bill: an agent that looped all weekend, a key shared by the whole team, no way to know which project burned the budget.

Recommended agent: LiteLLM

💰 Free self-hosted (open source); you still pay for your model calls — but capped this time. Level: Advanced 🌍 United States Free open-source version (MIT license) to host yourself; Enterprise offer with support by quote

Price verified on 2026-10-29.

LiteLLM (litellm.ai), an open-source proxy that sits between your applications and the model providers (OpenAI, Anthropic, Google…). Everything goes through it: it hands out virtual keys per person, project or agent, tracks every expense and cuts things off cleanly when a budget is reached — instead of discovering the bill at the end of the month.

How it works

Trigger: One of your agents or applications calls a model — checks the key and its remaining budget, routes the call to the right model and accounts for the spend, and you get spending attributed to the right key, and calls blocked as soon as a cap is crossed.

Step-by-step setup

Install LiteLLM self-hosted (Docker) with its database — essential: without it, budgets aren't enforced.
Plug your model provider accounts (your OpenAI, Anthropic keys…) into the proxy.
Create one virtual key per application, person or agent, with a monthly budget and allowed models.
Point your agents at the proxy's address, and set an alert at 50% and 80% of each budget.

⚠️ Guardrails: The guardrail is only worth anything if it's enabled: LiteLLM's documentation points out that without a connected database, budgets simply aren't checked. Test the blocking with a small key before trusting the setup, and also keep the providers' native caps as a second line of defense.

🔧 Technical details

Installation: Self-hosted
Prerequisites: Python via uv/pip (SDK or “litellm[proxy]” proxy) or Docker; API keys for the targeted model providers; a database is required for the proxy server's features (virtual keys, spend tracking), according to the documentation.
Hardware: non_applicable
Self-hosting: yes — self-hosted AI Gateway (Python binary or Docker, AWS/GCP Terraform templates provided); a hosted proxy offer also exists from the publisher.
License: open_core — MIT License for the whole repo except the enterprise/ folder, which is under a separate Enterprise license (official LICENSE file); GitHub classifies the license as “Other (NOASSERTION)” for this reason.
API & integrations: API: yes — OpenAI-compatible API (/chat/completions, /responses, /embeddings, /images, /audio, /batches, /rerank, /messages, etc.) exposed by the gateway. · SDK: yes — official Python SDK (litellm / litellm-core), plus a Rust core. · CLI: yes — litellm CLI to launch the proxy (e.g. litellm --model ...). · Webhooks: Not documented · MCP: yes — MCP Bridge (SDK) and MCP Gateway documented; support for the A2A protocol as well.
Models: More than 100 LLM providers in the OpenAI format (OpenAI, Anthropic, Gemini, Bedrock, Azure, Vertex AI, vLLM, NVIDIA NIM, etc.), with your own provider keys (BYOK); local models accessible via compatible providers (e.g. Ollama, vLLM).
Data: Requests pass through your gateway to the providers; logs and spend tracking are centralized in your deployment (built-in admin interface), according to the official documentation.
Documented limits: budgets and rate limits are features configurable per virtual key, not published technical caps.
Operations: Virtual keys, spend tracking, guardrails, load balancing, centralized logging and a built-in web admin interface in the gateway, according to the official documentation.

Official sources: docs.litellm.ai · github.com · github.com — verified on November 1, 2026.

Find your agent with the quiz →

Frequently asked questions

How much does LiteLLM cost?

Free self-hosted (open source); you still pay for your model calls — but capped this time. Free open source version (MIT license) to host yourself; Enterprise offer with support by quote. Price verified on October 29, 2026.

Is LiteLLM free?

Yes. Free open source version (MIT license) to host yourself; Enterprise offer with support by quote.

Is LiteLLM hard to get started with?

Level: Advanced. Setup takes 4 steps; the first: Install LiteLLM self-hosted (Docker) with its database — essential: without it, budgets are not enforced.

What limits and guardrails should I know about with LiteLLM?

A guardrail is only worth anything if it's enabled: LiteLLM's documentation makes clear that without a connected database, budgets simply aren't checked. Test the blocking with a small key before trusting the setup, and also keep the providers' native caps as a second line of defense.