Expose Velm to MCP-capable LLM clients (Claude and others) over the Model Context Protocol.
Framing: this is not a new data model - it is a second wire protocol on the SAME inbound layer already specced in #23 (records/tasks/actions) and #85 (service desk). Keep it lean: wrap existing capability, don't fork it.
Surfaces to expose:
Identity: MCP clients are LLMs, so this is the agents-as-users (#68) authN/authZ case directly: an agent client authenticates as an agent user, and MCP enforces the SAME authZ as any other surface. No separate permission model.
One real concern MCP adds vs REST: it is built for discovery - an LLM client must be able to see what it can do without docs. Tool naming/descriptions matter more than endpoints.
Design fork to settle: wrap the #23 action surface directly, OR expose a purpose-built set of agent-safe tools? Lean: wrap the same actions but be deliberate about which are safe for an autonomous agent - same parse-then-confirm discipline as the service desk (#85). Not every #23 endpoint is a safe raw MCP tool.
Category: c:api, c:agents. Depends: #23 (action surface), #68 (agent identity), #84 (catalog).
Expose Velm to MCP-capable LLM clients (Claude and others) over the Model Context Protocol.
Framing: this is not a new data model - it is a second wire protocol on the SAME inbound layer already specced in #23 (records/tasks/actions) and #85 (service desk). Keep it lean: wrap existing capability, don't fork it.
Surfaces to expose:
Identity: MCP clients are LLMs, so this is the agents-as-users (#68) authN/authZ case directly: an agent client authenticates as an agent user, and MCP enforces the SAME authZ as any other surface. No separate permission model.
One real concern MCP adds vs REST: it is built for discovery - an LLM client must be able to see what it can do without docs. Tool naming/descriptions matter more than endpoints.
Design fork to settle: wrap the #23 action surface directly, OR expose a purpose-built set of agent-safe tools? Lean: wrap the same actions but be deliberate about which are safe for an autonomous agent - same parse-then-confirm discipline as the service desk (#85). Not every #23 endpoint is a safe raw MCP tool.
Category: c:api, c:agents. Depends: #23 (action surface), #68 (agent identity), #84 (catalog).