Hook PromptForge into Cursor via the existing MCP server

.cursor/mcp.json

@@ -0,0 +1,17 @@
+{
+  "mcpServers": {
+    "promptforge": {
+      "command": "cargo",
+      "args": [
+        "run",
+        "--quiet",
+        "-p",
+        "promptforge-mcp-server",
+        "--",
+        "serve",
+        "--stdio",
+        "prompts.toml"
+      ]
+    }
+  }
+}

README.md

@@ -82,6 +82,16 @@ Interactive prompt work against an already-running gateway:
 cargo run -p promptforge-dev -- prompts/greet.md "world" --watch
 ```
 
+### Cursor
+
+This repo ships `.cursor/mcp.json`, which spawns `promptforge-mcp-server` over stdio using `prompts.toml`. For a local Ollama model, start the matching gateway (loopback password `whatever`, already set in both files):
+
+```bash
+cargo run -p promptforge-gateway -- serve gateway.ollama.toml
+```
+
+Enable the **promptforge** MCP server in Cursor Settings → MCP. The harness then has four tools: `list_prompts`, `run_prompt`, `check_run`, and `need_prompt`. Ask the agent to list prompts or to run `hello`.
+
 ![Gloves and sparks](images/banner-04.png)
 
 ## How it works

gateway.ollama.toml

@@ -0,0 +1,23 @@
+# Local Ollama backend. No vendor keys. Serve from the repo root:
+#   cargo run -p promptforge-gateway -- serve gateway.ollama.toml
+#
+# The loopback password is the literal `whatever`. The MCP server in
+# prompts.toml uses the same string. Point `upstream` at `ollama list`.
+
+[server]
+bind = "127.0.0.1:8081"
+api_key = "whatever"
+
+[[endpoint]]
+id = "ollama"
+protocol = "openai"
+base_url = "http://127.0.0.1:11434/v1"
+api_key = ""
+
+[[model]]
+name = "writer"
+description = "A model suited for careful analysis, coding, and general assistance"
+context = 8192
+thinking = "never"
+upstream = "openhermes:latest"
+endpoints = ["ollama"]

prompts.toml

@@ -16,16 +16,18 @@ api_key = "${PROMPTFORGE_MCP_SERVER_API_KEY}"
 [paths]
 prompts = "prompts"
 
-# The prompts run against the same gateway the CLI uses; it holds the vendor key.
+# The prompts run against the same gateway the CLI uses. The password is a
+# local loopback bearer, not a vendor key; it must match [server].api_key in
+# gateway.ollama.toml.
 [gateway]
 url = "http://127.0.0.1:8081/v1"
-api_key = "${PROMPTFORGE_GATEWAY_API_KEY}"
+api_key = "whatever"
 
-# Live tools: both enabled here to preserve the development profile's behavior.
-# A configuration with no [tools] section runs in a true sandbox.
+# Fetch is local. Search needs a gateway [tools.web_search] provider; the
+# Ollama profile does not ship one, so it stays off here.
 [tools]
 web_fetch = true
-web_search = true
+web_search = false
 
 # Everything written under prompts/ is available. `*` stops at a separator and
 # `**` crosses one, so `**/*.md` reaches a prompt in a subdirectory too. The