MCP Server

An MCP server built from AGNT5 components: a function, a workflow and an agent, published as tools that Claude, ChatGPT, Cursor, VS Code and Codex can call. Clients sign in with AGNT5, and every tool call runs as a durable AGNT5 run you can open in Studio.

Tool	Component	What it shows
`lookup_order`	`@function`	A tool that answers right away, shown in its own view
`refund_order`	`@workflow`	Durable multi-step work; a long call hands back a run handle
`audit_orders`	`@workflow`	A 10-minute run with steps and progress, followed in the run card or with `get_run`
`ask_support`	`Agent`	An LLM with its own tools, behind one MCP tool

Files


app.py                              worker entry point: imports the server
agnt5.yaml                          project config, with mcp.servers
src/agnt5_mcp_server/
  functions.py                      lookup_order, issue_refund, audit_order (sample data)
  workflows.py                      refund_order, audit_orders
  agents.py                         support_agent and its get_order tool
  server.py                         the "support" MCPServer
  views/order.html                  the built order view, committed
ui/                                 the order view's source (Vite, vite-plugin-singlefile)
tests/test_server.py                checks the published tools and the built view

A worker publishes only the MCP servers defined in modules it imports, so app.py imports server.py. Add a tool by writing the component and adding it to the server with add_function, add_workflow or add_agent.

The order view

Clients that render MCP Apps (ChatGPT, Claude, Cursor, VS Code) show lookup_order's result in the order view, your own UI. Clients that don't read the result as text. A view is one self-contained HTML file: ui/ builds it with Vite and vite-plugin-singlefile, and server.py ships it with support.add_view("order", path=...) and shows it with view=order_view. The view gets the tool's result (structuredContent and text) from the host through @modelcontextprotocol/ext-apps, and makes no network requests of its own.

The built src/agnt5_mcp_server/views/order.html is committed, so neither agnt5 dev nor a deploy needs Node. After changing ui/, rebuild it (Node 22+):

cd ui && npm install && npm run build

A view can be up to 2 MB. Other tools without a view of their own show the built-in AGNT5 run card while their run goes on.

Run it

agnt5 dev

Once the worker registers, agnt5 dev prints the server's dev URL. Only you can sign in to it. Connect Claude Code to it:

agnt5 mcp connect support --dev --client claude-code

Then in Claude Code run /mcp, choose support and sign in with AGNT5. Other clients: --client cursor, --client vscode or --client codex.

The agent needs OPENAI_API_KEY: copy .env.example to .env and set it. The functions and workflows need nothing.

A long-running call

audit_orders audits every order for about 10 minutes (pass minutes, 1 to 30): one step per order and check, each preceded by a ctx.progress report and followed by a durable ctx.sleep. server.py publishes it with mode="background", so a call hands back a run handle at once, and the run carries on without the client, through worker restarts.

Claude Code and Codex read the handle as text and follow the run with the built-in get_run tool (cancel_run stops it). Ask "audit all orders for 2 minutes", then "how is the audit going?": each get_run shows the run's status, latest progress and, once it finishes, the flagged orders.
claude.ai, ChatGPT, Cursor and VS Code render MCP Apps, so the call shows the built-in AGNT5 run card: live status, steps, a progress bar, the result and Cancel run. In claude.ai or ChatGPT, add the server's deployed URL (see below) as a custom connector and sign in with AGNT5.

Deploy it

agnt5 secrets set --name OPENAI_API_KEY --type api_key   # for the agent
agnt5 deploy --env production
agnt5 mcp connect support --env production --client claude-code

agnt5 deploy prints the server's URL, https://api.agnt5.com/mcp/<project>/<env>/support, and warns if a server named in agnt5.yaml was not published. Studio shows it under Build → MCP Servers, with its tools, keys and calls. If a client can't connect, agnt5 mcp check support --env production probes the server the way a client does and says which step fails.

Test

uv run --with pytest pytest