Agent Adapters

MCP Runtime includes two agent-side adapters that attach governed identity to MCP traffic without requiring the agent framework to know anything about grants, sessions, or policy:

• mcp-runtime adapter proxy exposes a local Streamable HTTP MCP endpoint and forwards requests to an MCP Runtime route.
• mcp-runtime adapter stdio exposes a stdio MCP server process and forwards each JSON-RPC message to the same MCP Runtime HTTP route.

Both adapters only present issued identity values. They do not create grants, evaluate policy, or bypass the gateway. Platform admins author MCPAccessGrant resources first — scaffold with mcp-runtime access grant init when helpful — and the platform API issues MCPAgentSession values through POST /api/runtime/adapter/sessions when the adapter starts with --server and --agent. The gateway is the enforcement point.

The adapter surface is intentionally limited to stdio and Streamable HTTP, the two standard MCP transports. There is no separate legacy HTTP+SSE adapter.

How the adapter gets its identity

There are three supported ways to give an adapter its humanID, agentID, teamID, and sessionID:

1. Platform-issued session (recommended). The adapter calls POST /api/runtime/adapter/sessions. The platform derives the principal from your mcp-runtime auth login token, picks a matching enabled MCPAccessGrant, writes (or reuses) an MCPAgentSession, and returns the identity values. Optional --auto-refresh renews the session before expiry without restarting the adapter.
2. Explicit flags / environment. --human-id, --agent-id, --session-id, --team-id (or the matching MCP_RUNTIME_* env vars). Useful for testing and for inheriting an externally-managed session.
3. Anonymous mode (stdio only): --anonymous skips identity entirely so the adapter can target public/read-only runtime routes. Only the methods listed in --anonymous-methods are forwarded.

Mixed configurations are supported: identity flags always override values returned by the platform-issued session, so a caller can pin a specific field (e.g. a long-lived --session-id for a test) while letting the platform fill in the rest. The override survives every auto-refresh tick.

Platform-issued sessions — quickstart

mcp-runtime auth login --api-url https://platform.example.com

mcp-runtime adapter stdio \
  --runtime-url https://mcp.example.com/workspace-assistant-mcp/mcp \
  --server workspace-assistant-mcp \
  --agent ticket-triage-agent \
  --auto-refresh

What this does on each invocation:

1. The CLI calls POST /api/runtime/adapter/sessions with {serverName, namespace?, agentID}. namespace defaults to the principal's primary namespace.
2. The platform derives humanID from Principal.Subject (fallback to Email) and teamID from the principal's membership in the namespace's team.
3. The platform lists enabled MCPAccessGrant resources in that namespace, filters those whose serverRef.name matches and whose subject equals the caller or is empty (wildcard), and picks the grant with the highest MaxTrust. Ties are broken by oldest creationTimestamp.
4. The platform looks up an existing MCPAgentSession with the deterministic name adapter-<sha256-prefix(humanID,agentID,teamID,serverName)> in the namespace. If one exists and is not revoked, has more than 30 s until expiry, and its policyVersion matches the selected grant's, it is reused. Otherwise a fresh MCPAgentSession is applied with a 1 h TTL (capped at 24 h).
5. The response carries name, humanID, agentID, teamID, consentedTrust, policyVersion, and absolute expiresAt. The adapter uses name as X-MCP-Agent-Session on every outbound request.
6. With --auto-refresh, a background goroutine renews the session ~5 min before expiresAt and atomically rotates the identity. In-flight requests continue with the previous identity; subsequent requests pick up the new one without a restart. Transient platform errors are logged to stderr; the previous identity stays in place until a refresh succeeds.

Required grant

A grant must exist before the platform will issue a session. Example:

apiVersion: mcpruntime.org/v1alpha1
kind: MCPAccessGrant
metadata:
  name: triage-grant
  namespace: mcp-servers
spec:
  serverRef:
    name: workspace-assistant-mcp
  subject:
    # Any of these may be empty to act as a wildcard for that field.
    humanID: support-lead
    agentID: ticket-triage-agent
    teamID: team-acme
  maxTrust: high
  allowedSideEffects:
    - read
  policyVersion: v1
  toolRules:
    - name: add
      decision: allow
      requiredTrust: low
    - name: upper
      decision: allow
      requiredTrust: low

If the principal does not match any enabled grant for the server, the adapter-session endpoint returns 403 and the adapter refuses to start.

Explicit-identity mode

When you already have an MCPAgentSession and don't want the platform to pick the grant for you (for example in a fixed CI environment), set everything explicitly:

export MCP_RUNTIME_URL=http://localhost:18080/workspace-assistant-mcp/mcp
export MCP_RUNTIME_HUMAN_ID=support-lead
export MCP_RUNTIME_AGENT_ID=ticket-triage-agent
export MCP_RUNTIME_SESSION_ID=sess-ticket-triage-agent

mcp-runtime adapter proxy

Environment variable	Required	Purpose
MCP_RUNTIME_URL	yes	Absolute Streamable HTTP MCP route.
MCP_RUNTIME_HUMAN_ID	yes¹	Human identity (X-MCP-Human-ID).
MCP_RUNTIME_AGENT_ID	yes¹	Agent identity (X-MCP-Agent-ID).
MCP_RUNTIME_TEAM_ID	no	Team identity (X-MCP-Team-ID) for team-scoped grants.
MCP_RUNTIME_SESSION_ID	yes¹	MCPAgentSession name (X-MCP-Agent-Session).
MCP_RUNTIME_HOST_HEADER	no	Override the Host header for host-based ingress.
MCP_RUNTIME_LISTEN_ADDR	proxy	Local listener; defaults to 127.0.0.1:8099.
MCP_RUNTIME_PROTOCOL_VERSION	no	MCP protocol header. Defaults to 2025-06-18; the negotiated result.protocolVersion from the runtime's initialize response overrides it for the rest of the process.
--no-xforwarded flag	proxy	Pass this flag to suppress X-Forwarded-* headers forwarded to the runtime. Defaults to enabled (headers are sent). There is no corresponding env var.
MCP_RUNTIME_REQUEST_TIMEOUT	no	Go duration for adapter→runtime calls. Defaults to unbounded.
MCP_RUNTIME_MAX_INBOUND_BYTES	proxy	Caps inbound JSON-RPC bodies; over-cap responds 413. Defaults to 16 MiB.
MCP_RUNTIME_AUTH_HEADER	no	Static Authorization header injected on every runtime request (e.g. Bearer …).
MCP_RUNTIME_TLS_CLIENT_CERT / _KEY	no	PEM client cert / key for mTLS to the runtime.
MCP_RUNTIME_TLS_CA_BUNDLE	no	PEM CA bundle replacing the system trust store.
MCP_RUNTIME_ANONYMOUS	stdio	true enables anonymous mode.
MCP_RUNTIME_ANONYMOUS_METHODS	stdio	CSV allowlist of methods in anonymous mode.
MCP_RUNTIME_TOOLS_CACHE_TTL	stdio	Caches tools/list responses for this duration (e.g. 30s). Anonymous mode bypasses the cache.
MCP_RUNTIME_LOG_LEVEL	no	info logs runtime 4xx denials to stderr.

¹ Required unless --server (platform-issued session) or --anonymous is in use. With --server, missing fields are populated from the issued response.

The adapters inject these headers on every forwarded request:

X-MCP-Human-ID:      <humanID>
X-MCP-Agent-ID:      <agentID>
X-MCP-Team-ID:       <teamID>            (omitted when empty)
X-MCP-Agent-Session: <sessionID>
Authorization:       <MCP_RUNTIME_AUTH_HEADER>   (when set)

Incoming spoofed values for the four governance headers are stripped before the upstream call. MCP protocol headers (Mcp-Protocol-Version, Mcp-Session-Id, content-type, accept) are preserved.

Anonymous mode (stdio)

For public/read-only routes — for example a catalog discovery endpoint — identity is unnecessary and the stdio shim can run anonymous:

mcp-runtime adapter stdio \
  --runtime-url https://mcp.example.com/public-catalog/mcp \
  --anonymous \
  --anonymous-methods initialize,notifications/initialized,ping,tools/list,resources/list,prompts/list

Anonymous methods default to the protocol handshake plus the three read-only discovery calls. Any method outside the allowlist is rejected with a JSON-RPC -32601 error before the request leaves the adapter, so an agent SDK cannot accidentally call tools/call against a public route.

The tools/list cache is bypassed in anonymous mode: different anonymous callers can see different responses depending on what the runtime exposes publicly, and there is no safe shared cache key.

Direct HTTP clients

When the agent framework supports Streamable HTTP MCP and custom headers, you can call the runtime directly without the adapter. Mint a session with the platform API once, then attach the returned identity on every request. Use a platform login token or user API key for this call; service-only setup keys do not carry a human subject and cannot mint adapter sessions.

import asyncio
import os
import httpx

from agents import Agent, Runner
from agents.mcp import MCPServerStreamableHttp

async def main() -> None:
    platform_token = os.environ["MCP_PLATFORM_API_TOKEN"]
    async with httpx.AsyncClient() as http:
        resp = await http.post(
            os.environ["MCP_PLATFORM_API_URL"].rstrip("/")
            + "/api/runtime/adapter/sessions",
            json={
                "serverName": "workspace-assistant-mcp",
                "agentID": "ticket-triage-agent",
            },
            headers={
                "Authorization": f"Bearer {platform_token}",
            },
        )
        resp.raise_for_status()
        session = resp.json()

    async with MCPServerStreamableHttp(
        name="workspace-assistant-mcp",
        params={
            "url": os.environ["MCP_RUNTIME_URL"],
            "headers": {
                "X-MCP-Human-ID": session["humanID"],
                "X-MCP-Agent-ID": session["agentID"],
                "X-MCP-Team-ID": session.get("teamID", ""),
                "X-MCP-Agent-Session": session["name"],
            },
        },
    ) as server:
        agent = Agent(
            name="Governed Agent",
            instructions="Use MCP tools when they help.",
            mcp_servers=[server],
        )
        print((await Runner.run(agent, "Add 2 and 3.")).final_output)

asyncio.run(main())

This is the only path that requires the consumer to know about the platform API. For framework code that cannot attach headers, prefer the proxy or stdio adapter described below.

HTTP proxy adapter

Use the proxy when a framework can speak Streamable HTTP MCP but cannot attach the governance headers itself.

mcp-runtime adapter proxy \
  --runtime-url https://mcp.example.com/workspace-assistant-mcp/mcp \
  --server workspace-assistant-mcp \
  --agent ticket-triage-agent \
  --auto-refresh

Then point the framework's MCP URL at the local proxy:

http://127.0.0.1:8099/mcp

The proxy forwards to the exact --runtime-url route. The local request path is accepted for client compatibility; it is not appended upstream. Query strings from the configured URL and client request are merged. By default the proxy adds X-Forwarded-* headers; set --no-xforwarded to suppress them on loopback paths where they only add audit noise.

The proxy also exposes:

• GET /healthz, GET /livez, GET /readyz — 204 No Content when running.
• GET /metrics — delegates to ProxyConfig.MetricsHandler when wired up (typically a Prometheus exporter backed by RuntimeTransport.Meter). Returns 404 when no metrics handler is configured.

Inbound JSON-RPC bodies over --max-inbound-bytes (default 16 MiB) get HTTP 413 with a JSON-RPC parse-error body so the agent SDK can recover.

This shape works for LangChain, LlamaIndex, CrewAI, custom Python/Go/Node services, or any other MCP-aware runtime that can talk to a Streamable HTTP URL.

Stdio shim

Use the shim when an IDE or client only launches stdio MCP commands (Cursor, Claude Desktop, similar):

{
  "mcpServers": {
    "workspace-assistant-mcp": {
      "command": "/absolute/path/to/bin/mcp-runtime",
      "args": [
        "adapter", "stdio",
        "--runtime-url", "https://mcp.example.com/workspace-assistant-mcp/mcp",
        "--server", "workspace-assistant-mcp",
        "--agent", "ticket-triage-agent",
        "--auto-refresh"
      ],
      "env": {
        "MCP_PLATFORM_API_URL": "https://platform.example.com",
        "MCP_PLATFORM_API_TOKEN": "..."
      }
    }
  }
}

The shim reads newline-delimited JSON-RPC messages from stdin, posts them to the runtime, and writes responses back to stdout.

Behaviour worth knowing:

• initialize is forwarded synchronously so the runtime Mcp-Session-Id is captured before later requests. The negotiated protocolVersion from the result body is also captured and used on subsequent calls.
• Streamable HTTP text/event-stream responses are streamed through frame by frame so server-to-client requests and progress messages flow without waiting for the runtime to close the response. The shim watches each SSE frame for notifications/tools/list_changed and invalidates its tools/list cache when it sees one.
• The shim leaves MCP_RUNTIME_REQUEST_TIMEOUT unset by default so long-running tool calls are not cut off. Set it when fail-fast behaviour is preferable.
• Platform 4xx denials (e.g. trust_too_low) are returned to stdio clients as JSON-RPC errors. Runtime denials matching session_expired / session_not_found are repackaged with error.data.runtime_status = "session_expired" so the SDK can choose to re-initialize.
• Idempotent reads (tools/list, resources/list, prompts/list, ping) retry on 502/504/connection-reset with exponential backoff (100 ms → 200 ms → 1 s cap). tools/call never retries automatically.

Expected outcomes

• A low-trust allowed tool call succeeds when the grant, session, and tool rule permit it.
• If the active session consents to less trust than a tool requires, the runtime returns trust_too_low and the adapter surfaces it as a JSON-RPC error to the client.
• Disabling or revoking the platform-side grant/session blocks calls without changing adapter configuration. With --auto-refresh, the next refresh tick may detect the new state (no matching grant → 403, surfaced in the adapter's stderr; the previous identity remains in use until expiry).
• Restarting the adapter against a revoked or expired session yields a fresh MCPAgentSession automatically, since the reuse predicate excludes revoked/near-expiry sessions.