MCP, A2A, OAuth & Skills API

All endpoints on this page are served by the Core Proxy (:8100) — the Rust runtime plane.

Base URL	http://localhost:8100
Auth	Gateway JWT (X-Gateway-Authorization: Bearer <jwt>) or API key (X-API-Key / Authorization: Bearer sk_brutor_api_...), plus X-Tenant-ID where the tenant isn’t in the credential
Public (no auth)	A2A discovery well-knowns, /health
Errors	{"error": "<message>"} with HTTP status

Guides: MCP Clients, A2A Agents, OAuth, Agent Skills.

MCP gateway

Base: /v1/proxy/mcp/{server_id} — the governed front door to every registered MCP server. Speaks JSON-RPC 2.0 over POST plus an optional SSE listener (streamable HTTP transport).

Method	Path	Purpose
POST	/v1/proxy/mcp/{server_id}	JSON-RPC 2.0 request (initialize, tools/list, tools/call, resources/read, …)
GET	/v1/proxy/mcp/{server_id}	SSE listener — send Accept: text/event-stream (streamable HTTP)
ANY	/v1/proxy/mcp/{server_id}/{path}	Sub-path passthrough to the upstream server
GET	/v1/proxy/mcp/{server_id}/.well-known/oauth-protected-resource	OAuth discovery (gateway-rewritten)
GET	/v1/proxy/mcp/{server_id}/.well-known/openid-configuration	OIDC discovery (gateway-rewritten)
GET	/v1/proxy/mcp/health	MCP subsystem health

Protocol versions & dialects

The gateway is bilingual across the MCP spec’s two eras (the spec calls this “dual-era”):

• Handshake dialect (≤ 2025-11-25): initialize → Mcp-Session-Id → notifications/initialized. Fully supported, unchanged — every existing client and server keeps working.
• Stateless dialect (2026-07-28): no handshake or sessions; every request carries params._meta (io.modelcontextprotocol/protocolVersion, .../clientCapabilities) and the MCP-Protocol-Version / Mcp-Method / Mcp-Name HTTP headers; capability discovery via the server/discover method. Serving this dialect from the gateway’s synthesized servers (VMCP, Agent Skills, Web Search) is gated behind MCP_STATELESS_ENABLED (default off).

Two behaviors are always on, regardless of the flag:

• Header/body consistency enforcement — a request whose Mcp-Method or Mcp-Name header contradicts the JSON-RPC body is rejected with HTTP 400 + JSON-RPC -32020 (Header Mismatch). Absent headers (legacy clients) pass; lying headers never do. This closes the spoofing gap where header-routing infrastructure sees a different action than the body performs.
• Dialect discovery & recording — at capability discovery the gateway probes upstreams (server/discover first when the flag is on, initialize fallback) and records each server’s negotiated protocol_version and dialect (handshake | stateless) on the server record; the proxy speaks each upstream’s own dialect.

Error-code note: the 2026-07-28 spec retires -32002 (resource not found) in favor of -32602; the gateway accepts both from upstreams during the transition.

Tool call example

curl http://localhost:8100/v1/proxy/mcp/hubspot-server \
  -H "Authorization: Bearer sk_brutor_api_..." \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/call",
    "params": {
      "name": "get_contacts",
      "arguments": { "limit": 10 }
    }
  }'

Responses are standard JSON-RPC 2.0: {"jsonrpc": "2.0", "id": 1, "result": {...}} or {"jsonrpc": "2.0", "id": 1, "error": {...}}.

Approvals

Tool calls subject to human-in-the-loop approval are blocked until approved; retry with the one-time X-Approval-Token header. See Approvals and the Portal API approvals endpoints.

Virtual MCP (VMCP)

Base: /v1/proxy/vmcp

Method	Path	Purpose
POST	/v1/proxy/vmcp/{virtual_server_id}	JSON-RPC to a virtual server aggregating multiple MCP servers
GET	/v1/proxy/vmcp/health	Health

A virtual server presents one endpoint exposing the tools of many upstream servers. Tool names are namespaced by server slug; every tools/call is delegated to the originating upstream, so per-server OAuth works transparently — no extra client work.

OAuth proxy

Base: /v1/proxy/oauth/{mcp_server_id}/{idp_server_id} — brokered OAuth for MCP servers behind identity providers.

Method	Path	Purpose
GET	/.../.well-known/oauth-protected-resource	Discovery metadata with gateway-rewritten URLs
GET	/.../.well-known/openid-configuration	OIDC discovery, gateway-rewritten
POST	/.../token	Authorization-code → token exchange
POST	/.../revoke	Revoke
GET	/.../tokens	List the caller’s stored tokens
POST	/.../register	RFC 7591 Dynamic Client Registration

Real tokens never leave the gateway

On POST /token, the real upstream OAuth access token is encrypted and stored server-side; the client receives a Gateway JWT instead. Subsequent MCP calls through the gateway attach the real token upstream automatically.

A2A v1.0 (spec-compliant)

Base: /a2a/{tenant_id}/{card_name} — Agent-to-Agent protocol v1.0 endpoints for agents published through the gateway.

Method	Path	Purpose
POST	/a2a/{tenant_id}/{card_name}/message:send	Synchronous message send
POST	/a2a/{tenant_id}/{card_name}/message:stream	Streaming message (SSE)
GET	/a2a/{tenant_id}/{card_name}/tasks	List tasks
GET	/a2a/{tenant_id}/{card_name}/tasks/{task_id}	Get task
POST	/a2a/{tenant_id}/{card_name}/tasks/{task_id}:cancel	Cancel task
GET	/a2a/{tenant_id}/{card_name}/tasks/{task_id}:subscribe	Subscribe to task updates (SSE)
POST	/a2a/{tenant_id}/{card_name}/tasks/{task_id}/pushNotificationConfigs	Create push-notification config
GET	/a2a/{tenant_id}/{card_name}/tasks/{task_id}/pushNotificationConfigs	List push configs
GET	/a2a/{tenant_id}/{card_name}/tasks/{task_id}/pushNotificationConfigs/{config_id}	Get push config
DELETE	/a2a/{tenant_id}/{card_name}/tasks/{task_id}/pushNotificationConfigs/{config_id}	Delete push config
GET	/a2a/{tenant_id}/{card_name}/extendedAgentCard	Extended (authenticated) agent card

Discovery (unauthenticated)

Method	Path	Purpose
GET	/.well-known/agent-card/{tenant_id}/{card_name}	Public agent card (served as application/json)
GET	/.well-known/a2a-jwks/{tenant_id}	Tenant Ed25519 JWK Set (RFC 7517) for verifying card signatures

Agent cards are signed with the tenant’s Ed25519 key; verifiers resolve the signature kid via the JWKS endpoint.

curl http://localhost:8100/.well-known/agent-card/default/revenue-sentinel

Legacy A2A v0.3 dialect

Method	Path	Purpose
POST	/v1/proxy/a2a/outbound	Governed outbound call to an external agent
POST	/v1/proxy/a2a/inbound	Governed inbound call from an external agent

Delegation context is carried on an HMAC-signed x-brutor-delegation-* header chain. Prefer the v1.0 surface for new integrations.

Agent Skills

Skills are served by the per-tenant system MCP server — there is no separate skills protocol. Guide: Agent Skills.

Endpoint	POST /v1/proxy/mcp/system-agent-skill-server-{tenant} (e.g. system-agent-skill-server-default)
Protocol	MCP JSON-RPC 2.0 (initialize, tools/list, tools/call)

initialize returns the protocol preamble plus the caller’s access-scoped skill catalog in instructions. tools/list advertises exactly five generic verbs (skills themselves are not tools):

Tool	Layer	Arguments (required in bold)
skills__list	L1 discover	—
skills__load	L2 load	skill_name
skills__read_resource	L3 act	skill_name, path
skills__run_script	L3 act	skill_name, script, args, approval_token
skills__render_template	L3 act	skill_name, template, vars

• skills__list → {"skills": [{"name", "description"}, ...]} (access-scoped).
• skills__load → two content blocks: protocol preamble + unmodified SKILL.md, then {"skill": {name, version, description}, "manifest": {resources, scripts, templates}}. Executes nothing; audited as skill_load.
• skills__run_script → the script’s output only (sandbox execution), or a {"execution_id", "status": "pending_approval", ...} envelope when approval-gated — retry with approval_token. Audited as skill_run.
• skills__read_resource / skills__render_template → governed file content / rendered text; audited as skill_read / skill_render.
• Failures return isError: true in the tool result (MCP transport is always HTTP 200); the governance verdict lands on the audit row.

A legacy skill__<name> tools/call is tolerated as an alias for skills__load; it is unadvertised and never executes.

Deprecated REST lifecycle

The former REST surface — POST /v1/proxy/skills/execute, GET /v1/proxy/skills/executions/{id}, POST .../cancel, GET /v1/proxy/skills/catalog — is deprecated. execute still answers during the migration window (with a deprecation warning) but will be removed; migrate to the MCP surface above.

Skill Runner sandbox

Skills execute in the Skill Runner sandbox — internal-only on :8210, JWT-authenticated with SKILL_RUNNER_JWT_SECRET (must match the Core Proxy), read-only rootfs, all capabilities dropped. It is never exposed publicly.

Access control

• MCP servers and skills are resource-group gated — bind them to the caller’s resource group in the Control Plane (Admin API). Skill access control is resource-group-only.
• Group per request resolves via X-Resource-Group header → primary group → first membership.
• Guardrails apply on mcp_input/mcp_output, skill_input/skill_output, and a2a_inbound/a2a_outbound surfaces when enabled — see Guardrails.