agentgateway

agentgateway is an open-source data plane for conventional service traffic and AI-native protocols. Its official homepage describes routing and governance for HTTP, gRPC, TCP, major LLM providers, self-hosted inference, Model Context Protocol traffic, and agent-to-agent communication. The appeal is consolidation: a platform team can apply routing, policy, observability, and cost controls without operating a separate gateway for each traffic type. It remains infrastructure software, so production value depends on deployment design, identity integration, policies, monitoring, and safe operations.

Capabilities and workflow fit

The concrete feature set is: 1. Native MCP and agent-to-agent protocol support. 2. Routing for hosted LLMs and self-hosted inference. 3. HTTP, gRPC, and TCP service traffic support. 4. MCP authentication, authorization, and tool scoping. 5. Traffic observability, auditing, and cost controls. 6. Standalone and Kubernetes deployment paths. Evaluate these as parts of a workflow, not isolated checklist items. Start with one representative task and document the input, expected result, permissions, human approval point, timeout behavior, and recovery path. Run ordinary cases plus missing data, revoked access, malformed input, and upstream outages. Track successful completion, correction time, latency, and cost per accepted result. That evidence is more useful than a polished demonstration.

For a system that reads company data or takes actions, review least-privilege access, credential storage, encryption, retention, deletion, audit logs, subprocessors, regional processing, and whether customer data is used for model training. Confirm export options and what happens when a subscription ends. Keep a qualified person responsible for consequential output until the workflow demonstrates reliable production behavior.

Pricing and purchasing checks

The homepage calls agentgateway open source, and the source describes a $0 software-license tier with separate self-hosting costs. No hosted-plan pricing was published at /pricing. Budget for compute, logging, high availability, upgrades, on-call ownership, and any commercial support agreement. Ask the vendor to put currency, taxes, billing interval, renewal terms, included usage, overages, support, and cancellation terms in writing. A pilot should include a spending cap and realistic peak volume. Usage-priced products can accumulate retries and agent-loop charges. Open-source software still carries infrastructure and operating costs.

Pros, limitations, and alternatives

The strongest advantages are unifies ai-native and conventional traffic; open-source code supports self-managed deployment; provides protocol-aware controls for mcp and a2a; can centralize routing and audit policy. Important limitations are self-hosting shifts availability and upgrades to the operator; a broad gateway can become critical infrastructure; policy design requires careful least-privilege work; no public hosted-plan price was found. Each limitation needs an owner and a test. Permission risk calls for scoped accounts and audit review; reliability risk calls for repeatable failure tests; uncertain pricing calls for a dated quote based on measured workload.

Relevant comparisons include Cloudflare AI Gateway and Anthropic MCP, plus Composio and Flowise. They are comparison points for adjacent integration, automation, observability, document, transcription, or infrastructure approaches, not guaranteed direct substitutes. Compare candidates on the same data and acceptance criteria. Do not infer feature parity from a category label.

Practical use cases and verdict

The clearest uses are 1. Authenticate and authorize MCP client traffic. 2. Route model requests across providers. 3. Audit tool calls made by internal agents. 4. Consolidate service, inference, MCP, and A2A traffic. Begin with the narrowest case that has a known baseline for time, errors, and cost. Define an acceptance threshold before implementation, such as target completion rate, maximum review time, and maximum cost per completed workflow. Preserve a manual fallback and test behavior when access expires or an integration changes.

agentgateway belongs on a shortlist when its verified capabilities remove concrete steps from one of those workflows. It should not be selected solely for broad feature counts or an AI label. Production adoption should follow only after current commercial terms, security controls, output quality, failure recovery, and exportability have been checked with the buyer’s own users and data.