Generate a scoped token and connect any MCP client — Claude Code, Claude Desktop — to your project endpoint. Reads return the full graph; writes go through validated tools.
Your architecture graph, the scoped MCP endpoint, and a Claude agent — reads and writes flowing both ways. Three steps from token to a building agent.
Create a read or read+write bearer token for your project — sha256-hashed at rest and revocable anytime.
Point Claude Code, Claude Desktop, or any MCP client at your project endpoint with the token.
Reads return the full graph; writes go through validated, audited tools — no separate brief required.
The MCP endpoint runs server-side and re-derives the token user's org role on every call — it never bypasses your access model.
Per-user bearer tokens (read or read+write), sha256-hashed at rest, revocable at any time.
add / connect / arrange / remove components and connections are schema-checked and scoped to your membership and role.
Every agent mutation is written to the activity log with source=mcp — fully auditable, same as a human edit.
A focused toolset that mirrors the app's own service layer — nothing an agent can do that a member couldn't do in the UI.
list_orgs · list_projects · get_project · get_architecture · search_library — full context, no scraping.
create_project · update_project · delete_project (owner) — scaffold and manage projects without leaving your agent.
add_container · add_component · update_component · remove_component — build the graph node by typed node.
connect_components · remove_connection · arrange_architecture — wire and lay out the system.
list_skills · create_skill · update_skill · delete_skill — manage the org's reusable Copilot skills.
The same graph the Copilot reads is the graph agents extend — one source of truth, no separate brief.