agentgateway
An API gateway supporting the MCP protocol, with built-in mcpAuthorization authorization policies and mcpGuardrails protection policies. It uses CEL expressions for access control and can implement request interception and response modification through the external ExtMCP gRPC service.
Timeline (last 90 days)
agentgateway's Ollama provider goes through the OpenAI-compatible endpoint, which completely ignores Ollama's native "think": false parameter, causing reasoning models' chain of thought to consume the entire max_tokens budget, the content to return an empty string, finish_reason to be length, and the HTTP status code to still be 200.
agentgateway v1.6.0 is a unified gateway that exposes an OpenAI-compatible /v1/chat/completions endpoint externally and can mount multiple backend models.
In agentgateway's testing, qwen3.8 (27.3B, Q4_K_M quantization, 262K context) had a valid answer rate of 90.0% (54/60), an accuracy rate among answered questions of 98.1% (53/54), and an overall accuracy rate of 88.3% (53/60).
In agentgateway's testing, nemotron-3.5-lightning (32.9B sparse MoE, 1M context) had a valid answer rate of 45.0% (27/60), an accuracy rate among answered questions of 100% (27/27), and an overall accuracy rate of 45.0% (27/60).
In agentgateway's testing, claude-sonnet-5 had a valid answer rate of 98.3% (59/60), an accuracy rate among answered questions of 93.2% (55/59), and an overall accuracy rate of 91.7% (55/60), with a latency p50 of 1080ms, p95 of 1412ms, and a cost of $0.000704 per call.
agentgateway has a built-in MCP authorization policy mcpAuthorization, which evaluates mcp.tool.name, mcp.tool.target, and JWT claims using CEL rules.
In agentgateway, mcp.tool.arguments is marked as post-request and is only populated for access logs after the call has executed, whereas mcpAuthorization rules are evaluated before the call, when the arguments are not yet populated.
On agentgateway v1.5.0, the CEL evaluation result of every read_text_file call is exactly the same regardless of the path; tool-name rules work while path rules have no data to read.
ExtMCP is an external gRPC service that agentgateway calls before a tools/call request reaches the upstream server, and calls once more after the response returns.
agentgateway v1.6.0通过一条只允许get_forecast的授权规则，在tools/list中隐藏了其他工具，并拒绝了对format_address的调用。
All Facts (10)
agentgateway's Ollama provider goes through the OpenAI-compatible endpoint, which completely ignores Ollama's native "think": false parameter, causing reasoning models' chain of thought to consume the entire max_tokens budget, the content to return an empty string, finish_reason to be length, and the HTTP status code to still be 200.
50%Unverifiedagentgateway v1.6.0 is a unified gateway that exposes an OpenAI-compatible /v1/chat/completions endpoint externally and can mount multiple backend models.
50%UnverifiedIn agentgateway's testing, qwen3.8 (27.3B, Q4_K_M quantization, 262K context) had a valid answer rate of 90.0% (54/60), an accuracy rate among answered questions of 98.1% (53/54), and an overall accuracy rate of 88.3% (53/60).
50%UnverifiedIn agentgateway's testing, nemotron-3.5-lightning (32.9B sparse MoE, 1M context) had a valid answer rate of 45.0% (27/60), an accuracy rate among answered questions of 100% (27/27), and an overall accuracy rate of 45.0% (27/60).
50%UnverifiedIn agentgateway's testing, claude-sonnet-5 had a valid answer rate of 98.3% (59/60), an accuracy rate among answered questions of 93.2% (55/59), and an overall accuracy rate of 91.7% (55/60), with a latency p50 of 1080ms, p95 of 1412ms, and a cost of $0.000704 per call.
50%Unverifiedagentgateway has a built-in MCP authorization policy mcpAuthorization, which evaluates mcp.tool.name, mcp.tool.target, and JWT claims using CEL rules.
50%UnverifiedIn agentgateway, mcp.tool.arguments is marked as post-request and is only populated for access logs after the call has executed, whereas mcpAuthorization rules are evaluated before the call, when the arguments are not yet populated.
50%UnverifiedOn agentgateway v1.5.0, the CEL evaluation result of every read_text_file call is exactly the same regardless of the path; tool-name rules work while path rules have no data to read.
50%UnverifiedExtMCP is an external gRPC service that agentgateway calls before a tools/call request reaches the upstream server, and calls once more after the response returns.
50%Unverifiedagentgateway v1.6.0通过一条只允许get_forecast的授权规则，在tools/list中隐藏了其他工具，并拒绝了对format_address的调用。
50%