agentgateway
agentgateway.dev
- Free plan
- Open source
- Windows
- macOS
- Linux
- Web
- API
- Self-hosted
What agentgateway is
agentgateway is a free, open-source HTTP and gRPC gateway for routing, securing, observing, and governing traffic among services, language models, MCP servers, and agents. Its LLM gateway provides an OpenAI-compatible API for providers including OpenAI, Anthropic, Bedrock, Gemini, Vertex, and OSS Llama runs. Self-hosted inference routing supports vLLM, TGI, and Triton, with latency-, cost-, and model-aware routing. MCP capabilities include server discovery, scoped tool calls, and audit trails; A2A routing connects runtimes such as LangChain, CrewAI, and ADK. Security options include JWT and OIDC, API keys, authorization policies, TLS and mTLS with certificate rotation, and rate limiting. Guardrails check MCP calls and responses for prompt injection and unsafe payloads, while prompt guards support checks and redaction. Observability includes OpenTelemetry, Prometheus token-usage histograms, Jaeger traces, and request logs with latency, token counts, and realized USD cost. It can run as a standalone binary, in Docker, or on Kubernetes.
Who it suits
It suits tool-builders, platform engineers, and AI enthusiasts working with service, LLM, MCP, or agent-to-agent traffic. Platform teams can use its listed security, observability, and spending controls.
Good to know, and what to watch
- Free open-source gateway with self-hosted deployment.
- Routes LLM traffic through an OpenAI-compatible API.
- Supports MCP discovery, scoped calls, and audit trails.
- Includes per-key or per-team token and dollar caps.
- Offers TLS automation, canary routing, and rate limiting.
- Deployment is self-hosted.
- Listed commercial support is partner-provided.
Plans
Every plan and priceagentgateway (open source)
Free plan
Getting started
- Open https://agentgateway.dev/
- Choose a standalone binary, Docker, or Kubernetes deployment
- Run it in a self-hosted Linux environment
- Connect model providers through its OpenAI-compatible API
- Configure authentication, authorization, TLS, rate limits, and routing policies
Limits to check
The listed open-source plan is free and self-hosted. It is provided as an open-source gateway deployed by the user.
The verdict
agentgateway combines routing, security, observability, and cost controls for service and AI-related traffic in a self-hosted gateway. Its open-source plan is free; commercial support is listed through a partner.
Questions people ask
Is agentgateway free?
Yes. The listed agentgateway open-source plan costs nothing.
Which platforms does it support?
Its listed platforms are API and Linux, with a self-hosted deployment model.
How can it be deployed?
The maker describes a standalone binary, Docker, or Kubernetes deployment.
Which model providers integrate with it?
Listed integrations include OpenAI, Anthropic, Gemini, Bedrock, Azure OpenAI, Mistral, DeepSeek, Ollama, Vertex AI, xAI, and Snowflake.
Does it support MCP and agent-to-agent traffic?
Yes. It provides MCP discovery, scoped tool calls, audit trails, guardrails, and A2A routing between agent runtimes.
What the maker publishes
- A2A support
- The A2A gateway supports capability discovery, modality negotiation and long-running task collaboration between agents. source
- API and configuration
- Standalone deployments use flat YAML configuration, while Kubernetes deployments use a controller, Gateway API resources and documented API references. source
- Authentication
- Documented authentication methods include JWT, API keys, basic authentication and MCP authentication specifications. source
- Authorization
- Fine-grained authorization uses role-based policies and the CEL expression engine. source
- Binary installation
- The official installer is a curl script that downloads a release binary, verifies its checksum and installs it into /usr/local/bin. source
- Company origin
- Solo.io originally created agentgateway and donated it to the Linux Foundation in August 2025. source
- Cost controls
- Features include scoped virtual keys, per-key or per-team token and dollar spend caps, and a dashboard grouped by model, provider, and user. source
- Cost controls
- Virtual keys, token and dollar caps, model cost catalogs and dashboards attribute and enforce agent spending. source
- Deployment
- The maker describes deployment options as a standalone binary, Docker, or Kubernetes. source
- Deployment
- The project provides standalone binary, Docker and Kubernetes deployment paths. source
- Documentation
- Documentation is split into standalone and Kubernetes sites with quickstarts, integration guides, configuration references and API documentation. source
- Enterprise support
- Solo Enterprise for agentgateway advertises 24×7 coverage, upstream collaboration and operational guidance. source
- Guardrails
- MCP guardrails inspect calls and responses for prompt injection and unsafe payloads, while prompt guards support checks and redaction. source
- Guardrails
- Guardrails include regex checks, OpenAI moderation, AWS Bedrock Guardrails, Google Model Armor and custom webhooks. source
- Inference
- Self-hosted inference routing supports vLLM, TGI, and Triton, with latency-aware, cost-aware, and model-aware routing. source
- Inference routing
- Kubernetes Inference Gateway extensions route self-hosted models using GPU utilization, KV-cache use, LoRA adapters and work-queue depth. source
- Integrations
- The site lists model integrations including OpenAI, Anthropic, Gemini, Bedrock, Azure OpenAI, Mistral, DeepSeek, Ollama, Vertex AI, xAI, and Snowflake. source
- Integrations
- Official integration categories cover identity providers, AWS, Google Cloud, Azure, LLM providers, MCP, networking tools, agent frameworks and NetBird. source
- Intended users
- The site invites tool-builders, platform engineers, and AI enthusiasts to join the community and describes the controls as suited to platform teams. source
- Licence
- The open-source project is licensed under Apache-2.0. source
- Limit: fast-changing configuration
- An independent review cautioned that the project is actively developing and its configuration surface is not yet stable. source
- Limit: insecure defaults
- A hands-on review found optional API-key authentication, fail-open deny expressions and an unauthenticated localhost admin interface as defaults requiring hardening. source
- Limit: operational complexity
- A reviewer said the Kubernetes controller and policy model add too much machinery for a single service with one model provider and a fixed tool list. source
- LLM gateway
- It offers one OpenAI-compatible interface for providers including OpenAI, Anthropic, Gemini and Bedrock, with budgets, failover and load balancing. source
Checked against the maker's pages on 19 Oct 2026.