Apache-2.0 agentic proxy that can expose stdio, HTTP, SSE, and Streamable HTTP MCP servers through a managed gateway with federation, OAuth/JWT authentication, RBAC/CEL policy, CORS, TLS, observability, and Kubernetes Gateway API support.
Agentgateway can aggregate and expose many downstream MCP targets through one endpoint; every target's permissions become part of the gateway surface., Stdio targets are spawned by the gateway process, so commands such as package runners or containers inherit the gateway host's trust boundary., Demo configs use permissive CORS and local authorization servers; lock down origins, headers, issuers, audiences, JWKS, and resource metadata before production use., CEL/RBAC policies can inspect MCP tool names, arguments, results, prompts, resources, request bodies, JWT claims, API keys, and backend metadata; test policies before relying on them., Remote MCP proxying, OAuth provider adaptation, Kubernetes routing, TLS termination, and guardrails are infrastructure changes that should go through normal security review.
Privacy notes
MCP requests, tool arguments, tool results, prompt names, resource names, session IDs, JWT/API-key claims, raw request/response bodies, logs, traces, and telemetry may pass through the gateway., CEL policy and observability features may buffer or inspect request and response bodies depending on configuration., OAuth/JWT metadata, bearer tokens, API keys, DCR secrets, Keycloak/Auth0/Okta settings, and upstream MCP credentials must be protected as secrets., Any data returned by downstream MCP tools can still be sent onward by the MCP client to the configured model provider.
Author
agentgateway
Submitted by
oktofeesh1
Claim status
unclaimed
Last verified
2026-06-06
Decision playbook
Review trust signals before you adopt
Signals are present but mixed. Use the checklist below to confirm the source and operational safety for your environment.
Compare context
Selected
0
Current score
63
Baseline
—
Delta
No baseline selected
No major trust-signal divergence detected in the current selection.
Source and provenance checks
Needs review
Confirm ownership and provenance before trusting install instructions.
Source link availableRequired
Open the canonical repository and verify ownership.
Done
Source provenance statusRequired
Marked as source-backed.
Done
Metadata reviewed
No reviewed flag detected in metadata.
Pending
Safety and privacy checks
Complete
Validate risk disclosures before installation or API wiring.
Safety notes presentRequired
Review the listed safety guidance before running commands.
Done
Privacy notes presentRequired
Review data handling notes before connecting accounts or secrets.
Done
Trust level risk gateRequired
Trust level does not block evaluation.
Done
Package and install checks
Needs review
Check package metadata and artifact integrity signals.
Install payload available
Install or copy payload is available for review.
Done
Package verification flag
No package verification flag provided.
Pending
Checksum metadata
No checksum provided for downloaded artifact.
Pending
Compare-driven decision checks
Needs review
Use compare context to validate trade-offs before adoption.
Compare tray has multiple entries
Add at least one more entry to compare trust differences.
5 safety and 4 privacy notes across 5 risk areas. Review closely: credentials & tokens, permissions & scopes, network access, third-party handling.
5 areas
SafetyPermissions & scopesAgentgateway can aggregate and expose many downstream MCP targets through one endpoint; every target's permissions become part of the gateway surface.
SafetyExecution & processesStdio targets are spawned by the gateway process, so commands such as package runners or containers inherit the gateway host's trust boundary.
SafetyPermissions & scopesDemo configs use permissive CORS and local authorization servers; lock down origins, headers, issuers, audiences, JWKS, and resource metadata before production use.
SafetyCredentials & tokensCEL/RBAC policies can inspect MCP tool names, arguments, results, prompts, resources, request bodies, JWT claims, API keys, and backend metadata; test policies before relying on them.
SafetyCredentials & tokensRemote MCP proxying, OAuth provider adaptation, Kubernetes routing, TLS termination, and guardrails are infrastructure changes that should go through normal security review.
PrivacyCredentials & tokensMCP requests, tool arguments, tool results, prompt names, resource names, session IDs, JWT/API-key claims, raw request/response bodies, logs, traces, and telemetry may pass through the gateway.
PrivacyNetwork accessCEL policy and observability features may buffer or inspect request and response bodies depending on configuration.
PrivacyCredentials & tokensOAuth/JWT metadata, bearer tokens, API keys, DCR secrets, Keycloak/Auth0/Okta settings, and upstream MCP credentials must be protected as secrets.
PrivacyThird-party handlingAny data returned by downstream MCP tools can still be sent onward by the MCP client to the configured model provider.
Safety notes
Agentgateway can aggregate and expose many downstream MCP targets through one endpoint; every target's permissions become part of the gateway surface.
Stdio targets are spawned by the gateway process, so commands such as package runners or containers inherit the gateway host's trust boundary.
Demo configs use permissive CORS and local authorization servers; lock down origins, headers, issuers, audiences, JWKS, and resource metadata before production use.
CEL/RBAC policies can inspect MCP tool names, arguments, results, prompts, resources, request bodies, JWT claims, API keys, and backend metadata; test policies before relying on them.
Remote MCP proxying, OAuth provider adaptation, Kubernetes routing, TLS termination, and guardrails are infrastructure changes that should go through normal security review.
Privacy notes
MCP requests, tool arguments, tool results, prompt names, resource names, session IDs, JWT/API-key claims, raw request/response bodies, logs, traces, and telemetry may pass through the gateway.
CEL policy and observability features may buffer or inspect request and response bodies depending on configuration.
OAuth/JWT metadata, bearer tokens, API keys, DCR secrets, Keycloak/Auth0/Okta settings, and upstream MCP credentials must be protected as secrets.
Any data returned by downstream MCP tools can still be sent onward by the MCP client to the configured model provider.
Prerequisites
Agentgateway binary, container image, or source checkout from the upstream release/container/docs path.
At least one downstream MCP target, such as a stdio command, remote MCP server, SSE server, or Streamable HTTP server.
MCP client that can connect to the exposed Streamable HTTP or SSE route.
Auth, CORS, TLS, route, and policy configuration reviewed before shared or remote exposure.
Kubernetes Gateway API environment if using the Kubernetes controller path.
Schema details
Install type
cli
Troubleshooting
No
Source repository stats
Scope
Source repo
Collection metadata
Estimated setup
35 minutes
Difficulty
advanced
Full copyable content
docker pull ghcr.io/agentgateway/agentgateway
About this resource
Content
Agentgateway is an open-source agentic proxy for AI-native traffic, including
MCP, A2A, and LLM gateway use cases. Its MCP gateway mode can connect to
downstream MCP servers over stdio, HTTP, SSE, or Streamable HTTP, then expose
them through a managed gateway route for MCP clients.
The basic example shows a stdio MCP target being exposed through a gateway that
serves Streamable HTTP under /mcp and SSE under /sse. The MCP
authentication example shows protected MCP routes with OAuth resource metadata,
JWT validation, CORS policy, and provider adaptation for Keycloak, Auth0, and
Okta-style environments.
These sources were reviewed on 2026-06-06. Prefer the live repository,
README, docs, MCP docs, Kubernetes docs, basic example, MCP authentication
example, Streamable HTTP implementation, CEL schema, GitHub Releases, and GHCR
package page for current setup, transport, auth, policy, and deployment
behavior.
Features
Proxy stdio MCP servers and expose them over gateway-managed HTTP routes.
Proxy remote MCP servers and serve OAuth protected-resource metadata.
Aggregate multiple MCP targets behind one gateway endpoint.
Serve MCP traffic over Streamable HTTP and SSE.
Add MCP authentication with JWT validation, JWKS, resource metadata,
Keycloak/Auth0/Okta provider adaptation, and DCR-related endpoints.
Apply CORS, TLS, API key, JWT, RBAC, and CEL policy controls.
Inspect MCP request context such as tool names, arguments, results, prompts,
resources, method names, and session IDs for policy and logging.
Deploy standalone or with the Kubernetes controller and Gateway API support.
Installation
Install from the upstream release or container path. The published container
package can be pulled with:
docker pull ghcr.io/agentgateway/agentgateway
Configure a reviewed MCP backend in agentgateway, start the gateway, and connect
your MCP client to the exposed route:
Follow the upstream quickstart, MCP docs, basic example, and MCP authentication
example for the current config schema, route paths, auth settings, and
Kubernetes deployment path.
Use Cases
Put one governed gateway in front of several local or remote MCP servers.
Convert a stdio-only MCP server into a Streamable HTTP or SSE endpoint for
clients that cannot spawn the process directly.
Add OAuth/JWT protection and resource metadata to remote MCP access.
Apply CEL/RBAC policy to tool calls, prompts, resources, and request context.
Run MCP gateway routing as part of a Kubernetes Gateway API deployment.
Safety and Privacy
Agentgateway sits on the trust boundary between clients and tools. Review every
target, route, CORS rule, auth issuer, audience, JWKS URL, resource metadata
field, TLS setting, and CEL policy before exposing it outside a local lab.
Stdio targets can spawn commands on the gateway host, and aggregated targets can
make a single endpoint much more powerful than any individual server appears.
Observability and policy inspection are useful, but they can expose sensitive
request details. MCP tool arguments, results, prompts, resource names, session
IDs, request/response bodies, JWT claims, API keys, and upstream credentials may
be visible to logs, traces, policy expressions, or model providers through the
MCP client.
Duplicate Check
No agentgateway/agentgateway entry, Agentgateway MCP Gateway entry,
agentgateway MCP proxy entry, or matching source URL was found in content/mcp.
Show that Agentgateway MCP Gateway is listed on HeyClaude. Paste this Markdown into your README — it renders the badge and links back to this page.
[](https://heyclau.de/entry/mcp/agentgateway-mcp-gateway)
How it compares
Agentgateway MCP Gateway side by side with 3 alternatives on trust, install, platform support, and disclosed safety notes — all from reviewed registry metadata.
Apache-2.0 agentic proxy that can expose stdio, HTTP, SSE, and Streamable HTTP MCP servers through a managed gateway with federation, OAuth/JWT authentication, RBAC/CEL policy, CORS, TLS, observability, and Kubernetes Gateway API support.
MIT-licensed MCP proxy server that aggregates multiple downstream MCP servers behind one HTTP service, exposing each configured server through SSE or Streamable HTTP endpoints with auth tokens and tool filtering.
✓Agentgateway can aggregate and expose many downstream MCP targets through one endpoint; every target's permissions become part of the gateway surface.
Stdio targets are spawned by the gateway process, so commands such as package runners or containers inherit the gateway host's trust boundary.
Demo configs use permissive CORS and local authorization servers; lock down origins, headers, issuers, audiences, JWKS, and resource metadata before production use.
CEL/RBAC policies can inspect MCP tool names, arguments, results, prompts, resources, request bodies, JWT claims, API keys, and backend metadata; test policies before relying on them.
Remote MCP proxying, OAuth provider adaptation, Kubernetes routing, TLS termination, and guardrails are infrastructure changes that should go through normal security review.
✓mcpo exposes proxied MCP tools as HTTP endpoints, so every tool permission from the wrapped MCP server becomes reachable through the OpenAPI surface.
Use a strong API key and avoid publishing the server to untrusted networks without authentication, TLS, and network controls.
Proxied tools may read, write, delete, send, deploy, or execute depending on the wrapped MCP server.
Config-file and hot-reload mode can add or change proxied servers while mcpo is running.
OAuth and custom header support can store or forward access tokens, bearer tokens, and service credentials.
✓tbxark MCP Proxy Server can spawn stdio subprocess MCP servers and connect to remote SSE or Streamable HTTP MCP endpoints.
Every exposed route inherits the permissions of its downstream server; use per-server `authTokens` and tool filters instead of exposing all tools broadly.
The docs support remote config URLs and an `--insecure` TLS flag; avoid untrusted configs and do not disable TLS verification outside controlled testing.
Token-in-route fallback can expose credentials through browser history, proxy logs, metrics, and referrers; prefer authorization headers when clients support them.
Logging can be enabled globally or per server, so review logs before routing secrets, prompts, tool arguments, or tool outputs.
✓Unla can expose existing MCP servers and REST APIs as new MCP endpoints, so the final tool surface depends on every imported config.
Generated endpoints may proxy read/write API operations; review tool descriptions and permissions before connecting an autonomous agent.
The upstream README notes that Unla is under rapid development and that documentation can lag behind implementation.
Replace default admin credentials and secrets before exposing the management UI or gateway outside a trusted development network.
Hot-loaded gateway configuration, multi-tenant routing, OAuth pre-authentication, and management UI access should be governed like production API infrastructure.
Privacy notes
✓MCP requests, tool arguments, tool results, prompt names, resource names, session IDs, JWT/API-key claims, raw request/response bodies, logs, traces, and telemetry may pass through the gateway.
CEL policy and observability features may buffer or inspect request and response bodies depending on configuration.
OAuth/JWT metadata, bearer tokens, API keys, DCR secrets, Keycloak/Auth0/Okta settings, and upstream MCP credentials must be protected as secrets.
Any data returned by downstream MCP tools can still be sent onward by the MCP client to the configured model provider.
✓Tool names, OpenAPI schemas, prompts, arguments, tool outputs, proxied server URLs, headers, API keys, OAuth tokens, logs, and generated docs may be visible to mcpo, connected clients, and model providers.
OpenAPI docs can reveal the available tool surface and parameter shapes for sensitive MCP servers.
Store config files and OAuth token storage in locations with appropriate filesystem permissions.
Avoid proxying private MCP tools into shared OpenAPI workspaces unless all downstream users are authorized for the underlying tools.
✓The proxy may process downstream server commands, environment variables, headers, auth tokens, remote config URLs, tool names, tool arguments, tool outputs, prompts, resources, logs, and route keys.
Stdio subprocesses inherit configured environment variables and can access whatever local paths, network services, or credentials those commands can reach.
Remote config files can reveal internal MCP server names, endpoint URLs, auth headers, tool filters, and secret placeholders.
Keep config files, tokens, `.env` values, logs, and Docker volume mounts out of public repositories and shared support output.
✓Unla may receive prompts, tool names, tool arguments, REST API parameters, upstream MCP responses, generated endpoint metadata, logs, admin credentials, JWT secrets, OAuth tokens, and configuration history.
Imported REST API and MCP server configs can include private upstream URLs, headers, tokens, tenant identifiers, and service credentials.
Persisted config backends and container logs may retain sensitive endpoint definitions and operational metadata.
Do not commit real gateway configs, API credentials, admin passwords, JWT secrets, OAuth client secrets, or exported tenant data.
Prerequisites
Agentgateway binary, container image, or source checkout from the upstream release/container/docs path.
At least one downstream MCP target, such as a stdio command, remote MCP server, SSE server, or Streamable HTTP server.
MCP client that can connect to the exposed Streamable HTTP or SSE route.
Auth, CORS, TLS, route, and policy configuration reviewed before shared or remote exposure.
Python 3.11 or newer, with uv recommended by the upstream project.
An MCP server command, SSE endpoint, or Streamable HTTP endpoint you are authorized to proxy.
API key or other access control planned before exposing the HTTP server beyond local testing.
Network binding, reverse proxy, root path, and TLS settings reviewed before deployment.
Go toolchain for source install, or Docker for the GHCR image deployment path.
A v2 JSON configuration file defining `mcpProxy` and downstream `mcpServers`.
Downstream MCP server commands, URLs, headers, environment variables, and secrets prepared before proxy startup.
Public base URL, bind address, route type, auth token policy, logging policy, and tool filters reviewed before exposing the proxy beyond localhost.
Docker available for the all-in-one deployment, or Kubernetes and Helm if using the packaged cluster deployment.
Administrator username, administrator password, and JWT secret replaced before first start.
At least one existing MCP server or REST API configuration ready to import into the gateway.
Network, TLS, reverse proxy, and port exposure reviewed before sharing generated MCP endpoints.