Claude Agent on Workbench

Claude Agent is the Workbench operator surface: in-app /assistant, Campaign Management private chat turns, and external Claude Code / Agent SDK attach. It is not a ninth product agent and does not replace EcoGentic customer chat.

Branding: use Claude Agent (not “Claude Code Agent”) per Anthropic guidelines.

Usage map

Consumer	Needs from Workbench	Does not need
Campaign Management	Private `/claude-agent/` with `tool_profile: campaign_management` (CM MCP tool pack + read-only Workbench subset); `/campaigns/_generate/`; campaigns CRUD/trigger; JWT	Workbench mutating tools (hard-denied on this profile); Claude CLI in the CM image
Workbench `/assistant`	Ready Agent SDK + default LLM + in-process MCP	CM
External Claude Code / Agent SDK	`ANTHROPIC_BASE_URL` → Workbench host `:8001` (`POST /v1/messages`); optional dual-auth `POST /mcp`	Workbench in-image `_bundled/claude` (client supplies its own SDK/CLI)

Packaging

The Claude Agent SDK bundles the Claude binary (_bundled/claude). Workbench Docker uses python:3.14-slim (Debian/glibc) so the platform wheel includes that binary. Alpine/musl images are unsupported for the in-app Operator.

No Anthropic API key in the image.
Model traffic always uses Workbench POST /v1/messages → default LLM (LLMService.get_config("default")).
Readiness: GET /api/v1/claude-agent/status → ready: true when enabled + default LLM + SDK + resolved CLI.

Catalog agents

Admin Agents list product agents (spend/money personality, interaction science, intelligent sales, real-time recommender, conversational, EcoGentic, etc.) plus Claude Operator. Operator reaches them via in-process MCP tools and call_endpoint on private /api/v1/* paths.

Prefer:


GET  /api/v1/campaigns
POST /api/v1/algorithms/money-personality/process   (confirm=true)
POST /api/v1/algorithms/sentimental-equilibrium/process

/public/v1/... paths are accepted and remapped where needed. See MCP Support for Workbench MCP details.

Campaign Management

Set:


# CM side — reach the WB2 agent runtime
WORKBENCH_PRIVATE_API_BASE_URL=http://ecosystem-workbench2:8001/api/v1
CM_AGENT_MCP_API_KEY=<optional; hardens CM /agentic/mcp>
 
# WB2 side — mount the CM tool pack for tool_profile: campaign_management
AGENT_APP_CM_MCP_URL=http://ecosystem-campaign-management:54310/api/extensions/campaign-platform/agentic/mcp
AGENT_APP_CM_MCP_API_KEY=<same value as CM_AGENT_MCP_API_KEY>

CM bridges chat turns to private Claude with tool_profile: campaign_management (flag CAMPAIGN_PLATFORM_AGENTIC_WB2_RUNTIME). The WB2 runtime plans the turn and calls back into CM’s /agentic/mcp tool pack with the acting user’s JWT — campaign mutations apply through CM’s own intent pipeline (undo checkpoints + audit), and the profile hard-denies Workbench mutations. Generative analyze/execute and campaign ops use the same private API base with the user JWT.

External attach

Surface	Auth	Purpose
`POST /v1/messages`	Public path; model remapped to default LLM	What external Agent SDK / Claude Code calls when `ANTHROPIC_BASE_URL` = Workbench
`POST /mcp`	`Authorization: Bearer` + `X-API-Key`	Full HTTP MCP tool set

Copy live snippets from Workbench Administration → Agents → Access.

Ontology-aware Q&A

When Ontology Management is enabled, Operator MCP includes:

Tool	Role
`list_ontologies` / `describe_ontology`	Discover ontologies and enabled mappings
`resolve_ontology_term`	Map business terms → properties / fields
`generate_ontology_query`	Build a read-only Mongo pipeline or Trino SQL from Accepted mappings
`query_data_via_ontology`	Execute (`mode=auto\|filter\|aggregate`) against the mapped source

Example: “Using xyz ontology, how many customers spent money last month?” → resolve ontology by name → generate query → execute → report tool counts only.

Workbench Ontology Overview Test opens the agent drawer with that prompt prepared (composer filled, not auto-sent).

Related: MCP Support (Operator in-process vs HTTP /mcp vs stdio SDK).