
Agent网关横评:对比Bifrost、Kong、AgentCore、Gravitee等8款产品
来源标题:Agent Gateways Compared: Bifrost, Kong, AgentCore, Gravitee and 4 More
文章对比了Bifrost、Kong、AgentCore、Gravitee等8款agent网关产品,核心差异在于能否统一覆盖agent产生的三类流量:模型调用、MCP工具调用和A2A agent间通信。文章指出,选型关键看三点:是否看得全这三类流量、权限能否在agent到工具再到agent的跳转中保持隔离、以及单个任务能否在日志中读作一条完整链路。
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
Your support agent calls a model, pulls an invoice through an MCP tool, then hands the refund to a billing agent another team owns. A week later finance asks why that refund went out, and you have three disconnected stories: a model call in one dashboard, a tool call in another, and an agent-to-agent request that never passed through anything you control.
That gap is what an agent gateway exists to close. Picking one in 2026 comes down to three things: whether it sees all three kinds of traffic an agent creates, whether permissions stay separate as a task hops from agent to tool to agent, and whether one task still reads as one story in the logs.
An agent does not make one kind of call. It talks to a model over a provider API, to tools over the Model Context Protocol, and increasingly to other agents over Agent2Agent (A2A), the protocol the Linux Foundation took on in 2025.
An LLM gateway only sees the first of those. An MCP gateway sees the second. An agent gateway puts all three behind one front door: it authenticates the caller, decides which models, tools and agents that caller may reach, and records every hop of the task under the same identity.
The difference shows up the moment agents call each other. A2A has its own discovery file, the agent card, which tells a client where to send requests. If the card still points at the agent’s real address, clients will happily go around your gateway. Maxim’s explainer on agent gateway vs LLM gateway covers where the two categories overlap, if you are still deciding which layer you need.
Most of these map to OWASP’s excessive agency risk, which gets harder to contain with every agent you add to the chain. A gateway worth deploying closes all five.
How I evaluated them
I looked at five things that decide whether an agent gateway survives a security review and a real multi-agent rollout:
Everything below comes from each vendor’s own documentation, release notes and pricing pages, checked in November 2026.
- Bifrost
Bifrost is the open-source AI gateway from Maxim AI, written in Go and licensed Apache 2.0. It already handled model traffic and MCP tools; its Agent Gateway adds the third path, so A2A calls between agents go through the same deployment and the same virtual keys.
You register an existing A2A agent by its full agent card URL, either in the dashboard or in config:
{ "agents": [{ "name": "library-research", "agent_card_url": "https://research.example.com/.well-known/agent-card.json", "enabled": true, "allow_by_default": false, "virtual_key_ids": ["YOUR_VIRTUAL_KEY_ID"], "forward_accepted_credential": true }] }
Callers then talk to the agent through Bifrost, with their own key:
curl http://localhost:8080/agents/a2a/library-research/json-rpc
-H "x-bf-vk: $BIFROST_VK"
-H "Content-Type: application/json"
-H "A2A-Version: 1.0"
-d '{"jsonrpc":"2.0","id":"r1","method":"SendMessage",
"params":{"message":{"messageId":"q1","role":"user",
"parts":[{"text":"How do I reuse one HTTP client?"}]}}}'
What stands out
Best for: Bifrost is built for enterprises running mission-critical AI agents that need best-in-class performance, security and governance across every model, tool and agent call. It unifies the LLM gateway, MCP gateway and agents gateway into a single platform, and supports air-gapped, VPC and on-prem deployments for regulated industries, with full control over data, access and execution.
Pricing: open source is free. Enterprise is custom, with a 14-day trial.
Kong AI Gateway extends Kong’s API gateway to model, MCP and A2A traffic. Version 2.0, generally available since September 1, 2026, organizes AI traffic around entities (AI Models, AI MCP Servers, AI Agents and AI Policies) managed from a dedicated Konnect control plane, with on-prem data planes configured through decK.
An A2A agent is a few lines of declarative config (excerpt from Kong’s docs):
ai_gateway_agents:
- ref: kongair-flight-booking-agent name: kongair-flight-booking-agent type: a2a access: acls: allow:
- internal-teams config: url: https://booking-agent.internal.kongair.com route: paths:
- "/kongair-flight-booking"
What stands out
Best for: organizations already running Kong for API management that want agents under the same team, tooling and policies.
Pricing: a Konnect subscription; agent plugins need the AI Gateway Enterprise tier. Prices are not published.
AgentCore Gateway is AWS’s managed entry point for agents. It turns OpenAPI specs, Smithy models, Lambda functions and existing MCP servers into one virtual MCP server, and adds inference targets for model calls.
Best for: teams building agents on AWS that want a managed gateway wired into IAM, CloudWatch and Bedrock Guardrails.
Pricing: $0.005 per 1,000 invocations, plus separate charges for tool search, tool indexing and policy checks.
Google Cloud Agent Gateway is the policy enforcement point for agents on Gemini Enterprise Agent Platform. It governs MCP and A2A interactions with IAM policies and screens traffic with Model Armor. It became generally available in June 2026.
Best for: teams standardized on Gemini Enterprise Agent Platform and Google Cloud identity.
Pricing: billed on Agent Compute vCPU-hours, with Model Armor and logging billed separately.
Azure API Management adds AI gateway features to Microsoft’s API platform: token limits, semantic caching and content safety for model APIs, MCP servers built from REST APIs or proxied from existing ones, and imported A2A agent APIs.
<llm-token-limit counter-key="@(context.Request.IpAddress)" tokens-per-minute="5000" estimate-prompt-tokens="false" remaining-tokens-variable-name="remainingTokens" />
REST to MCP. An existing REST API in APIM can be exposed as an MCP server, secured with Entra ID OAuth.
A2A import, generally available. JSON-RPC A2A agents can be imported as APIs since June 2026; the gateway rewrites the agent card’s host and emits OpenTelemetry agent attributes.
Best for: Microsoft-centric enterprises that already publish APIs through API Management and authenticate with Entra ID.
Pricing: included in API Management tiers; the self-hosted gateway needs Developer or Premium.
Gravitee’s Gamma platform, first released on June 26, 2026, runs LLM, MCP and A2A proxies on one AI Gateway runtime. All three share one authentication chain, one policy chain and one observability chain, with OpenTelemetry tracing across every agent hop.
The LLM Proxy accepts both OpenAI and Anthropic request formats on the same proxy, so existing SDKs point at it unchanged.
Best for: API platform teams that want agent identity and authorization modeled alongside their existing APIs.
Pricing: the API Management core is open source; Agent Management starts at $2,500 a month on the Planet tier.
ContextForge is IBM’s open-source registry and proxy, licensed Apache 2.0. It federates MCP servers, REST and gRPC services, and A2A agents behind one endpoint. The latest release is v1.1.1, from November 10, 2026.
It installs with pip install mcp-contextforge-gateway or as a container, and an A2A agent is registered with one call:
curl -X POST "http://localhost:4444/a2a"
-H "Authorization: Bearer $MCPGATEWAY_BEARER_TOKEN"
-H "Content-Type: application/json"
-d '{
"name": "hello_world_agent",
"endpoint_url": "http://localhost:9999/",
"agent_type": "jsonrpc",
"passthrough_headers": ["Authorization", "X-Tenant-Id"],
"auth_type": "api_key",
"auth_value": "your-api-key"
}'
What stands out
Best for: teams that want a self-hosted, MCP-first gateway and are comfortable operating an open-source project themselves.
Agent Router is the project formerly called Envoy AI Gateway: same code, same maintainers, Apache 2.0, built on Envoy Gateway. It joined the Agentic AI Foundation as a standalone project on September 9, 2026.
Tool filtering is declared on the route. This MCPRoute backend only exposes issue-related tools from a GitHub MCP server:
backendRefs:
- name: github kind: Backend group: gateway.envoyproxy.io path: "/mcp/x/issues/readonly" toolSelector: includeRegex:
- .issues?.
What stands out
Best for: Kubernetes platform teams already running Envoy Gateway who want open-source model and MCP routing.
Pricing: free and open source; Tetrate sells an enterprise edition.
Start with the question that removes the most options: where do your agents live?
Then pressure-test the pick on the two things that bite later: does a key that can call an agent also quietly reach that agent’s tools, and can you rebuild one task from the logs alone?
One more gap worth naming: all eight govern traffic you route through them. Desktop agents can still reach MCP servers installed locally on a laptop. Bifrost Edge, currently in alpha, extends the same policies to employee devices by inventorying local MCP servers and enforcing allow or deny on the machine itself.
An agent gateway is a single control point between AI agents and everything they call: models, MCP tools and other agents. It authenticates the caller, decides which of those the caller may reach, and logs every hop so one task can be followed end to end. Maxim’s enterprise agent gateway guide walks through the architecture.
An MCP gateway governs tool calls only. An agent gateway adds model calls and A2A calls between agents under the same identity and policy. Bifrost, Kong, Gravitee and IBM ContextForge cover all three in one product, though their A2A support differs a lot in depth.
It depends on where your agents run. For agents spread across clouds and frameworks, Bifrost is the most complete of the eight, with native A2A over JSON-RPC, REST and gRPC, separate grants per path and one log session per task. For agents on one cloud’s platform, that cloud’s gateway is the simplest choice.
Once that agent calls tools or another agent on behalf of real users, usually yes. A single agent with three tools already means three sets of credentials and no shared log. A gateway gives you one place to revoke access and one record of what the agent did.
The agent card is how an A2A client learns where to send requests. If the gateway passes the card through unchanged, it still advertises the agent’s real address, and any client can skip the gateway. Bifrost, Kong and Azure API Management rewrite the card; Gravitee rewrites its endpoint URL; Agent Router serves it as-is. Maxim’s A2A agent card explainer shows what changes.
Yes. Bifrost, IBM ContextForge and Agent Router are all Apache 2.0. Bifrost covers models, MCP tools and A2A in one self-hosted gateway; ContextForge and Agent Router are MCP-first, with lighter A2A support. Kong and Gravitee have open-source cores, but their agent features sit behind commercial licenses.
There is no single best agent gateway. There is a best one for where your agents run and how much of their work you need to prove afterwards.
If every agent lives on one cloud’s platform, that cloud’s gateway is the obvious start. If they span clouds, frameworks and your own data center, the question is whether models, tools and agents share one policy and one log, or live in three systems you stitch together after an incident. Kong, Gravitee and Azure extend an API platform you may already run. ContextForge and Agent Router are solid open-source MCP gateways growing into A2A. Bifrost is the one that already governs all three paths under the same key, and Bifrost Edge carries those policies out to the laptop.
Whichever you pick, test it like an attacker would: give a key access to one agent, have that agent ask for a tool the key was never granted, and check that the logs show the whole task in one place.
Teams evaluating agent gateways can request a Bifrost demo or review the open-source repo.
This post was made in collaboration with the Bifrost team.
Short takes and discussions on X → https://x.com/sebuzdugan
Practical AI / ML videos on YouTube → https://www.youtube.com/@sebuzdugan/
Part
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。