MCPs

Several hosts expose MCP on the path /mcp. They are not interchangeable: a Workbench ewb_public_… key does not unlock Runtime scoring tools, and a modeling-server MCP OAuth token is not what MeWB accepts.

Workbench MCP

• Name: Workbench MCP (ecosystem-workbench)
• Live endpoints: POST https://mewbserver.ecosystem.ai/mcp and POST https://me.ecosystem.ai/mcp (both return the same unauthenticated 401). Local default: POST http://localhost:8001/mcp.
• Transport: Streamable HTTP. Docs describe MCP protocol 2026-07-28; the live server negotiated 2025-11-25 on initialize (11 Nov 2026, 13:16 SAST). Legacy SSE (GET /mcp/sse/, POST /mcp/messages/) was removed 26 October 2026.
• Auth: HTTP headers only, not a query parameter. One credential: X-API-Key: ewb_public_…, or Authorization: Bearer ewb_public_…, or Authorization: Bearer <Workbench JWT>. If Authorization already holds the public key, a JWT may be sent as X-User-Token. Mint keys under Administration → API Keys → Public API Access; a public key used alone on the gateway needs a linked internal API key. MeWB does not advertise MCP OAuth (well-known OAuth URLs on mewbserver also return the credential 401).
• Unauthenticated error (live): HTTP 401, Content-Type: application/json, body exactly {"detail":"Missing credential. Send Authorization: Bearer <jwt> or X-API-Key: ewb_public_..."}.
• What it does: Operator and catalog access to deployments, entities, enrichment, outreach, ontology, runtime testing, simulations, and call_endpoint into /api/v1 and /public/v1. Calls are logged to mcp_access_log.
• Live authenticated test (11 Nov 2026, 13:16 SAST): initialize HTTP 200, notifications/initialized HTTP 202, tools/list HTTP 200. Server: Ecosystem Workbench 1.0.0. tools/list returned 97 tools. Server instructions: with no catalog agent selected, every tool is available; the operator can list agents, use <agent id>, then clear (list_catalog_agents, select_catalog_agent, clear_catalog_agent); how-to questions use explain_catalog_agent. Resource mentioned by the server: ecosystem://agents.
• Documented resources (not separately listed in that tools/list response): ecosystem://deployments, ecosystem://deployment/{name}. Per-deployment snippets: GET /api/v1/deployments/{name}/mcp.
• Related: REST info (JWT) GET /api/v1/mcp/. In-process mode is the Ecosystem Agent at /assistant (session JWT, no public key). stdio harness: Python package ecosystem_workbench_mcp for Hermes / desktop. Cursor config uses a headers map on the /mcp URL.
• Health (open): GET https://mewbserver.ecosystem.ai/public/v1/health → {"status":"healthy","version":"1.0.0"}.
• Catalog UI: https://mewbserver.ecosystem.ai/ .

Live Workbench tools (97)

Grouped by family from live tools/list (11 Nov 2026, 13:16 SAST):

Catalog and environment: list_api_catalog, describe_environment, call_endpoint

Explain and recommend: recommend_algorithm, explain_algorithms, explain_convergence, explain_runtime, explain_simulation, explain_workbench, describe_workbench

Entities and records: entity, list_entities, count_records, explain_entity, search_product, get_record, get_related, inspect_entity, resolve_deep_link, update_current_form

Campaign channels: list_campaign_channel_catalog, get_campaign_channels, set_campaign_channel_selection, list_custom_channel_modules

Deployments and scoring: list_deployments, get_deployment_info, get_deployment_schema, score_customer, send_feedback

Enrichment: date_enrich, mcc_enrich, get_enrichment_job, personality_drift

Ontology and semantics: list_ontologies, describe_ontology, create_ontology, update_ontology, delete_ontology, create_ontology_from_data_source, list_ontology_mappings, suggest_ontology_mappings, suggest_ontology_mappings_multi, create_ontology_mapping, update_ontology_mapping, delete_ontology_mapping, resolve_ontology_term, generate_ontology_query, query_data_via_ontology, get_ontology_session_manifest, browse_semantics, resolve_semantics

Network selector: list_network_selector_configs, explain_network_selector, plan_network_selector, validate_network_selector, save_network_selector, fix_network_selector, assign_network_selector, unassign_network_selector, clone_network_selector, remove_network_selector, test_network_selector

Runtime testing: list_runtime_testing_suites, get_runtime_testing_suite, analyze_runtime_testing, plan_runtime_testing, start_runtime_testing_run, list_runtime_testing_runs, get_runtime_testing_run, cancel_runtime_testing_run, close_runtime_testing_run, compare_runtime_testing_runs, get_runtime_testing_artifact, delete_runtime_testing_suite

Simulations: start_simulation_run, cancel_simulation_run, get_simulation_run_status, get_simulation_logs, get_simulation_results, explain_simulation_result, get_simulation_config, upsert_simulation_config

Generate: check_generate_intent, analyze_generate, execute_generate, generate_solution_template, generate_project_setup_notebook

Outreach: list_outreach_templates, inspect_outreach_campaign, build_outreach_campaign, configure_outreach_campaign, run_outreach_campaign, get_outreach_run_status

Catalog agents: list_catalog_agents, select_catalog_agent, clear_catalog_agent, explain_catalog_agent

Runtime builtin MCP

• Name: Runtime builtin MCP (Java ecosystem-runtime, 0.9.7+)
• Endpoint: POST {RUNTIME_BASE}/mcp — typically port 8091. Example: POST http://localhost:8091/mcp. No hosted public Runtime MCP URL was found on me, meserver, mewbserver, or modelingserver. Clients must point at the customer’s Runtime host.
• Transport: JSON-RPC on POST /mcp. Send Accept: application/json, text/event-stream. Supports legacy (2024-11-05) and modern (2026-07-28, including server/discover). Per-request SSE is allowed; deprecated GET /mcp + POST /mcp/message still exists for old clients.
• Auth: Read-only tools (tools/list, listCampaigns, ontology validators, invoke, response) work without a key unless Spring --security is enabled. Mutating tools (campaignCreate, campaignDelete, generateClass, updateProperties*, learningTrigger) need Runtime MASTER_KEY as X-API-KEY or Authorization: Bearer {MASTER_KEY}. That key is not a Workbench ewb_public_… key.
• What it does: Score campaigns and close the learning loop: listCampaigns → validateInvokeRequest → invoke → validateResponseRequest → response. Also campaign admin, runtime ontology, and askAboutAlgorithms.
• Tools (docs): Scoring and campaigns — listCampaigns, invoke, response, updateProperties, updatePropertiesJson, generateClass, campaignCreate, campaignGet, campaignDelete, configGet, learningTrigger. Ontology — lookupRuntimeConcept, getClosedLoopRecipe, validateInvokeRequest, validateResponseRequest. Algorithms — askAboutAlgorithms. Resources include ecosystem://campaigns, ecosystem://config, ontology://runtime, ecosystem://algorithms/selection-guide.
• REST equivalents: POST {RUNTIME_BASE}/invocations, POST {RUNTIME_BASE}/response, plus /ontology/runtime catalog and validators.

Runtime MCP sidecar

• Name: Runtime MCP sidecar
• Image: ecosystemai/ecosystem-runtime-mcp (example tag arm64)
• Endpoint: FastAPI /mcp on the sidecar port (default 8092). Not the Java JSON-RPC server and not Workbench MCP. No public hosted sidecar URL was found on the four live hosts checked.
• What it does: MLFlow model download into Runtime (/refresh) and custom FastAPI routers compiled from a Workbench deployment API tab. OpenAPI at /docs.
• Auth / env: Configured with RUNTIME_URL, PORT, MLFLOW_TRACKING_URI, RUNTIME_CONFIG, MONGO_CONNECT. Auth follows that sidecar deployment; it is not the Workbench public-key scheme.

Modeling-server MCP (Data, AI, Modeling, Algorithms)

Host: https://modelingserver.ecosystem.ai/api/  (OpenAPI 3.1, platform version 0.7.9). Base path /api. The same path set exists on https://meserver.ecosystem.ai/  (/api/mcp/…), but meserver’s OAuth discovery currently advertises http://localhost:3002/api, which is not usable for external clients. Prefer modelingserver for OAuth metadata.

Unauthenticated calls to these MCP paths return HTTP 403 with an empty body (not MeWB’s JSON 401). OpenAPI marks a global JWT security scheme. MCP clients should use MCP OAuth (bearer token in the Authorization header) or a platform JWT. Scopes: mcp:tools, mcp:read, mcp:write. Bearer method: header only.

Live /info and /health were not readable without a token (403), so individual tool names were not listed from the running modeling MCP servers.

MCP OAuth (modeling / ecosystem-server)

This is not Workbench MCP auth.

• Discovery: https://modelingserver.ecosystem.ai/api/.well-known/oauth-authorization-server  (200)
• Protected resource: https://modelingserver.ecosystem.ai/api/.well-known/oauth-protected-resource  (200)
• Issuer / resource: https://modelingserver.ecosystem.ai/api
• Authorize: https://modelingserver.ecosystem.ai/api/oauth/authorize (OAuth 2.1 authorization code + PKCE S256)
• Token: https://modelingserver.ecosystem.ai/api/oauth/token (grants: authorization_code, refresh_token; client auth: client_secret_post, client_secret_basic, none)
• Register: https://modelingserver.ecosystem.ai/api/oauth/register (RFC 7591 dynamic client registration)
• Scopes: mcp:tools, mcp:read, mcp:write
• Token use: Authorization: Bearer <access_token> against the modeling MCP resources above
• Docs: https://modelingserver.ecosystem.ai/api/docs 

Domain Catalog MCP

• Endpoints: POST https://ontology.ecosystem.ai/catalog/mcp and POST https://ontology.ecosystem.ai/data/mcp
• Live check: GET returns HTTP 405 with Allow: POST and {"detail":"Method Not Allowed"}, so the POST MCP routes exist. Authenticated tools/list was not run in catalog verification. Earlier code review found no inbound auth on these routes (CORS only); treat that as unverified against a live POST.
• What it does: Domain Catalog / data ontology MCP used with the ontology download surface on ontology.ecosystem.ai. Related public (no JWT) modeling-server catalog APIs are tagged “Modeling: Domain Catalog” in the OpenAPI spec.

OpenAI-compatible model list and LLM gateway

GET /v1/models is OpenAI-compatible.

Related LLM endpoints on modelingserver (documented; POST not tested in catalog compile):

• POST https://modelingserver.ecosystem.ai/api/v1/chat/completions — OpenAI chat completions (streaming and non-streaming)
• POST https://modelingserver.ecosystem.ai/api/v1/responses — OpenAI Responses API compatibility
• POST https://modelingserver.ecosystem.ai/api/v1/messages — Anthropic Messages compatibility
• POST https://modelingserver.ecosystem.ai/api/v1/embeddings — OpenAI embeddings (forwarded to Ollama /api/embed)
• POST https://modelingserver.ecosystem.ai/api/modeling/llm/chat/completions — metered completions; docs say JWT and usage recording

On Workbench, set ANTHROPIC_BASE_URL to the Workbench host root so POST /v1/messages uses the Workbench default LLM.

Gaps (label explicitly)

• Authenticated Workbench initialize and tools/list succeeded on 11 Nov 2026 at 13:16 SAST against POST https://mewbserver.ecosystem.ai/mcp with a Workbench public key in the X-API-Key header: HTTP 200 / 202 / 200, 97 tools, server protocol 2025-11-25. POST https://me.ecosystem.ai/mcp was not re-tested with this key (it previously returned the same unauthenticated 401 as mewbserver).
• Modeling Data / AI / Algorithms / Modeling tool names were not live-listed (403 without a token).
• No public Runtime or sidecar MCP host is published on the four ecosystem.ai hosts checked.
• Campaign Outreach Agent and Network Selector Agent appear in the live product catalog; see Agents for operator context.