MCP server
AgentDesk runs a built-in Model Context Protocol server that exposes a project's shared memory to any AI client. This is how Claude Code, Codex, Gemini and web clients read and write the same project brain.
Transports
All clients connect to the hosted server over HTTP/SSE — GET /mcp/sse?project={name} and POST /mcp/messages, authenticated with an ad_ integration key. AgentDesk registers the server per project in a .mcp.json file, so each project exposes its own memory.
Tools
The server exposes tools across memory, history, work items, code and git. Read-only tools fetch context; mutating tools (marked ✎) change state.
Memory & context
| Tool | Purpose |
|---|---|
search_memory | Hybrid search across work items + events (semantic + keyword + relationship graph, rank-fused) |
record_event ✎ | Write a structured fact into memory |
pin_fact ✎ | Pin an explicit rule — project-wide or scoped to one code folder — always injected into context |
bootstrap_session | Load conventions, policies and components for a session |
get_session_tags / set_session_tags ✎ | Read / update the current session's phase, feature, bug ref |
History
| Tool | Purpose |
|---|---|
get_recent_history | Last N prompt/response pairs |
get_commits | Commits tagged by phase / feature / bug |
get_work_item_history | Linked prompts + commits for a work item |
get_file_history | Symbol-level changes for a file |
Work items
| Tool | Purpose |
|---|---|
create_entity ✎ | Create a draft work item |
list_work_items | Filter by type / status / due date |
get_item_by_number | Resolve BU0001 / FE0002 to full detail |
update_work_item ✎ | Change status or score |
record_work_log ✎ | Record delivery, verification and commit provenance against an item/use case |
run_work_item_pipeline ✎ | AI-summarise an approved item |
Code & git
| Tool | Purpose |
|---|---|
get_code_symbols | Cyclomatic complexity + hotspot scores |
get_db_schema | Full database schema dump |
commit_push ✎ | Stage, commit and push from any IDE |
Skills
| Tool | Purpose |
|---|---|
get_skill | Skill definition + project context |
run_skill ✎ | Execute a skill (same flow as the web UI) |
Example: Claude Code
Point Claude Code at the project's MCP server and it can search memory, read history and commit — all sharing the same brain as your other tools. A common first call is bootstrap_session (to load current conventions, facts and open work) followed by search_memory for whatever you're working on. At the end of implementation, agents can call record_work_log so the use case and generated OKF bundle show what changed.
Other MCP clients
Any MCP-capable agent can use the same server — Cursor, Cline, Roo Code, Windsurf, Zed, Continue.dev, Goose, OpenCode, Warp and Claude Desktop all connect with the standard mcpServers block; only the config-file location differs per client. The desktop app's External Skills & MCP view has a per-client setup card with a copy-paste config and the client's quirks.
Capture hooks (prompt, session and tool-observation logging) remain a Claude Code / Codex / Gemini feature; MCP-only clients get the full tool set without passive capture.
This server exposes a codebase's project memory. AgentData's MCP server exposes a database's semantic model. Different tools, same protocol.
Next: Shared prompts →