Worksona MCP Servers: Bringing 100+ Agents into Claude Desktop

The gap between a catalogue of well-defined AI agents and a working multi-agent environment has typically required custom integration code. The Worksona MCP Server closes that gap for Claude Desktop users: install once, and 100+ organizational role agents become available inside any Claude conversation as native tools.

What They Are

There are two related servers in this part of the portfolio. worksona-mcp-server is the full-featured TypeScript implementation, deployable either as a local STDIO server for Claude Desktop or as a Netlify-hosted HTTP service for team access. worksona-mcp-server-local is its locally-focused companion, optimized for individual developer use via the same STDIO transport.

Both implement the Model Context Protocol (MCP 2024-11-05) — the standard that Claude Desktop uses to discover and invoke external tools. The servers are protocol-native: MCP compliance is the foundation, not an adapter layer bolted on top of a proprietary API.

The agent roster spans 19 organizational categories: C-suite executives, board of directors, software architecture, security operations, data engineering, investor panels, and more. Each agent carries a role definition, domain expertise description, capabilities list, and a full system prompt.

Why It Matters

Three properties distinguish this approach from typical agent integration work.

Filesystem-as-registry. Adding an agent requires creating two files — a metadata.json and an agent.md — inside the teams/ directory. Removing an agent means deleting those files. The AgentRegistry discovers agents by globbing for metadata.json across all subdirectories. Git is the management system. There is no admin UI, no database migration, no API call to register a new agent.

Confidence-scored suggestions. The suggest_agents tool returns agents ranked by a confidence score, not a flat list. The score quantifies match quality, so a caller can decide programmatically whether to accept the top suggestion or surface all candidates for human review.

Four-pattern coordination with auto-selection. The WorkflowCoordinator encodes four meaningfully different multi-agent architectures. A suggestCoordinationType() function selects the appropriate pattern based on keyword analysis of the task description — reducing the cognitive overhead of multi-agent workflow design for routine tasks while allowing explicit override when needed.

How It Works

flowchart TD
    A[Claude Desktop] -->|STDIO / MCP| B[worksona-mcp-server]
    C[REST Client] -->|HTTP / MCP| B
    B --> D[AgentRegistry]
    D -->|glob metadata.json| E[teams/ directory]
    E --> F[100+ Agent Definitions\nmetadata.json + agent.md]
    B --> G[AgentExecutor]
    G -->|prompt assembly + LLM call| H[LLM Provider]
    B --> I[WorkflowCoordinator]
    I --> J{CoordinationType}
    J --> K[Sequential]
    J --> L[Parallel]
    J --> M[Hierarchical]
    J --> N[Consensus]
    B --> O[WorksonaAgentMemory]
    O -->|relevance-ranked| P[Session Context Store]
    B --> Q[Resources: 21 attachable]
    Q --> R[Task Templates\nWorkflow Definitions\nExample Prompts]

When Claude Desktop invokes a tool, the request arrives via STDIO transport. AgentRegistry has already discovered available agents from the filesystem. AgentExecutor assembles the appropriate system prompt, injects session memory from WorksonaAgentMemory, and makes the LLM call. Coordination workflows compose multiple executors using the selected pattern.

The same TypeScript codebase handles both STDIO and HTTP transports with no shared-state coupling — MCP tool definitions behave identically regardless of which transport carries the request.

For Claude Desktop setup, a single entry in claude_desktop_config.json pointing to the compiled dist/index.js and a WORKSONA_TEAMS_PATH environment variable is all that is required to activate the full agent roster.

Where It Fits in Worksona

The MCP servers occupy the protocol integration position in the portfolio. worksona-teams provides the agent content; the MCP servers make that content accessible to Claude Desktop. Where worksona-api is a standalone runtime for custom integrations, the MCP servers are specifically designed to extend what Claude can do natively inside its own interface.

This project anchors two patterns that appear across the portfolio: filesystem-as-registry and confidence-ranked agent suggestions. Both propagate to sibling MCP projects including agent-mcp and delegator-mcp.

The practical result is that an organization can maintain its entire agent library as a directory of text files in version control, and Claude Desktop users gain access to that library without any additional tooling.