Claude Setup
LiveConnect PMI to Claude in three flavours: Claude Desktop for personal analyst use, Cursor IDE for code-side integration, and the Claude API for production agents you build yourself. All three use the same MCP URL.
Pick your flavour
Same URL, same key. The difference is just where the conversation lives.
Claude Desktop
The Mac and Windows app. Best for personal analyst work, deal review, ad-hoc property questions. No code required.
Jump to setup ↓Option BCursor IDE
The AI-first code editor. Ideal if you do data analysis in notebooks or maintain Python integrations against the PMI API.
Jump to setup ↓Option CClaude API
For production agents. Slack bots, customer support tools, internal automation. Python and TypeScript SDKs supported.
Jump to setup ↓Option DOther Claude clients
Continue, Cline, Aider plugins, anything else MCP-compatible. Same URL, same Bearer header, same 28 tools.
Jump to setup ↓See it in action
If you've never connected an MCP server to Claude before, these public walkthroughs show what the experience looks like. PMI works the same way.
Connect your first MCP server
The canonical walkthrough. Same JSON config edit, same restart, same tool list, just our URL.
modelcontextprotocol.io ↑Anthropic · SupportClaude Desktop MCP guide
Anthropic’s own help-centre article. Step-by-step screenshots of the config file location and the connector panel.
support.claude.com ↑Option A: Claude Desktop
Claude's Mac and Windows app supports remote MCP servers natively. About 30 seconds to connect.
Open the config file
Easiest route: open Claude Desktop, go to Settings → Developer → Edit Config. That opens the file in your default editor and creates it if it doesn't exist.
Or open it directly:
Add the PMI block
If the file is empty, paste the whole thing. If you've already got other MCP servers, just add the pmi entry inside the existing mcpServers object.
Replace pmi_live_YOUR_KEY_HERE with your actual key from the dashboard.
Fully quit and reopen Claude Desktop
Cmd+Q on macOS, then relaunch. Closing the window isn't enough. The app needs a clean restart to load the new config.
Confirm it loaded
Click the plug icon at the bottom of the chat input box, then select Connectors. You should see “pmi” listed with 28 tools.
If you don't see it, jump to the troubleshooting section at the bottom of this page.
Try a real question
Open a new chat and try one of these:
Each tool call appears in the chat with an expandable JSON pane so you can verify the inputs and the response.
Option B: Cursor IDE
Cursor has first-class MCP support. Ideal if you're doing data analysis in notebooks, building a property dashboard, or maintaining integration code against PMI.
Open Cursor MCP settings
Three options, same end result:
- UI: Open Settings → Tools & MCP → New MCP Server.
- Project-level config: Create
.cursor/mcp.jsonin your project root. - Global config: Create
~/.cursor/mcp.jsonfor use across all projects.
Project-level config takes precedence if both exist.
Add the PMI block
Restart Cursor and verify
Open Settings → Tools & MCP. A green dot next to “pmi” means it's connected. Red means the connection failed (check the Authorization header, then check the URL has no trailing slash).
Use it in chat
Open Cursor's chat panel (Cmd/Ctrl+L) and try:
Option C: Claude API (Messages API)
For production agents. Slack bots, customer support flows, internal tooling, anything where you're calling Claude programmatically. Use the Anthropic Messages API with an mcp_servers parameter.
Python
TypeScript
What you get back
response.content is an array of typed blocks: text, mcp_tool_use, and mcp_tool_result. Iterate them in order to reconstruct what Claude did. Most apps just need the final text block.
Multi-turn conversations
To continue a conversation (the agent might want to call follow-up tools after seeing a result), feed response.content back as the assistant message and add a new user message:
Cost considerations
Each tool call is a separate Anthropic API turn. Anthropic charges for the model tokens used to formulate the call and parse the result. PMI bills separately based on which tools were called (same as REST). For a typical agentic interaction, expect 2 to 4 tool calls per user question.
Option D: Other Anthropic-powered clients
Anything that wraps Claude with MCP support (Continue, Cline, Aider plugins, etc.) uses the same URL and Bearer token. Look for an “HTTP MCP server” or “Remote MCP” config option and paste:
Best practices
- Don't put your master key in client configs. Create a sub-key per client (e.g.
pmi_live_claude_desktop_macos) so you can revoke it without breaking other integrations. - Set a credit cap on each sub-key. Limits how badly an agent gone wild can drain your allowance.
- Use
dry_run=truewhen prototyping. PMI tools accept this query param. The response shape mirrors the real call but no credits are spent. - Watch the connector panel in Claude Desktop. The tool log shows exactly which tools the model picked. If it consistently picks the wrong one, get in touch and we'll tune the description.
- Pair MCP with a short system prompt. Telling Claude things like “always quote the median, not the mean, when discussing rents” helps it format answers consistently.
Troubleshooting
Connector panel shows "pmi: 0 tools"
You're connected but the tools didn't enumerate. Almost always: the Authorization header isn't being passed. Re-check the JSON in claude_desktop_config.json, common typo is "Header" instead of "Authorization". Then quit and relaunch Claude Desktop fully.
Tool calls return "Authentication failed"
The token is being sent but PMI rejected it. Confirm the key starts with pmi_live_, that it's marked active in the dashboard, and that there's no leading or trailing space in the config.
Tool calls return "Monthly credit limit reached"
You’ve spent your allowance. Upgrade your tier in the dashboard, or wait until the next billing cycle starts.
Claude says "I tried to call pmi but the tool isn't available"
The MCP server isn’t connected. Click the connector panel; if "pmi" isn’t listed at all, the URL is wrong or the file isn’t being read. Re-do the config step.
Latency is bad on tool calls
First call from a fresh Claude Desktop session warms up the MCP connection, expect about 500ms to 1s overhead. Subsequent calls in the same session are sub-200ms beyond the underlying endpoint’s latency.
JSON config edits aren't taking effect
Common causes: invalid JSON (a trailing comma is the usual culprit), saved the wrong file, or didn't fully quit Claude Desktop. Validate your JSON with jq or paste it into an online JSON validator.