Connect Claude Code to plori with one prompt

Set up https://plori.ai/SKILL.md

https://plori.ai/SKILL.md is a short alias. The canonical address is https://plori.ai/.well-known/agent-skills/plori/SKILL.md, and both serve the same bytes. The skill is for the agent, not for you, so you do not need to read it. Its Connect section is the source for every step described here.

What does the agent do with that prompt?

In the verified run, a fresh Claude Code session received only this prompt. It did four things, then stopped:

1. Fetched the skill from the URL.
2. Ran claude mcp list and found no plori server configured.
3. Ran claude mcp add --transport http plori https://api.plori.ai/mcp.
4. Ran claude mcp list again to check the entry, then told the user that the tools load in the next Claude Code session or after typing /reload-plugins.

The skill tells the agent to reuse an existing plori server at that endpoint instead of adding a second one. The run added no duplicate entry.

How do you approve the connection?

Configuration alone does not authenticate the account. In a later interactive session, with the server already configured but not yet authenticated, the agent ran the pairing flow:

1. It called the plori MCP client's authenticate tool and kept the returned authorization URL unchanged.
2. It sent that URL once to https://api.plori.ai/oauth/pair.
3. It told the user: "Open https://plori.ai/device and enter code XXXX-XXXX". You can open that page on your phone.
4. It polled https://api.plori.ai/oauth/pair/poll about every 30 seconds while the user signed in with an email one-time code and approved the connection in the browser.
5. When the poll returned approved, the agent passed the returned callback_url to the client's complete_authentication tool.
6. It called list_agents, which answered with the one agent already on the account.

You only ever see the short address and the code. The authorization URL and the callback URL stay between the agent and the client, and the skill tells the agent not to ask you to paste either one into chat.

What did the real run prove?

We ran this on 7 September 2026 against production, using Claude Code 2.1.263 with Opus 5.

A fresh session given nothing but the prompt read the skill, configured the plori MCP server at https://api.plori.ai/mcp, and verified its own work with a second claude mcp list. No duplicate server entry appeared.

We exercised pairing in three rounds. The first round denied the first poll request under Claude Code's permission mode. The limits section below describes how that run continued. The second round expired after four minutes with expired_token because nobody approved the code, which is the correct outcome for an unattended pairing. In the third round the user approved in the browser, the poll returned approved, complete_authentication accepted the callback URL, and list_agents returned the account's one existing agent.

That last call is the whole proof. It shows an authenticated tool call reaching plori as the account owner. It does not create an agent, and it does not start billed work.

How do you check the connection safely?

1. Ask the agent to list its configured MCP servers, or run claude mcp list yourself. Look for plori pointing at https://api.plori.ai/mcp.
2. If the plori tools are missing from the current conversation, type /reload-plugins, or start Claude Code again. Saved configuration does not prove that the tools loaded.
3. Ask the agent to connect if it has not authenticated yet, and finish approval in the browser.
4. Ask it to list your plori agents. It calls list_agents, and the reply names the agents your account already has.
5. Stop there if all you wanted was the connection. Nothing in these five steps creates an agent or draws credits.

What should you run next?

The connection check is free. The first real job is not. When you want to see an agent do work, give your agent this second prompt:

Create a cloud agent on plori, send it a hello message, and show me its reply.

That one creates an agent and spends prepaid credits, because model usage and invoking an agent both draw on your balance. Run it when you are ready to pay for a run, not as part of the connection check. See pricing for the current plans and rates.

What are the limits?

Headless Claude Code cannot pair. Under claude -p, the client does not expose authenticate and complete_authentication for a server that still needs authentication, so the pairing flow has nothing to call. The skill instructs the agent to say so rather than keep trying. Add and authenticate the server from an interactive session first.

A permission denial ends the pairing path by design. Claude Code's default automatic permission mode denied the first poll request in one run. The skill's rule after any denial is to stop the pairing path immediately and show you the raw OAuth authorization URL instead. You open that URL, sign in, and approve there, and the client finishes on a localhost callback. On an explicit retry in the same run, Claude Code allowed the poll, so a denial changes the route to approval rather than blocking it.

Pairing lasts four minutes. That window fits inside the client's pending login. If it expires, the agent starts a fresh authentication rather than reusing the old authorization URL.

Common questions about connecting Claude Code to plori

Do I need to install a plugin or a skill first? No. The prompt points Claude Code at a URL it can read on its own. The skill says explicitly that installing a skill or plugin is optional, and the verified run installed neither.

Why did the plori tools not appear right after the agent added the server? Claude Code loads MCP tools when a session starts. In the verified run the agent said so and told the user what to do: type /reload-plugins in the current conversation, or start Claude Code again.

Does connecting cost anything? No. Adding the server, pairing, and calling list_agents spend no credits. Model usage and invoking an agent do draw on your prepaid balance. Current plans and rates are on pricing.

Can I do this in a headless run? Not the pairing part. claude -p does not expose the two authentication tools for a server that still needs authentication. Pair once in an interactive session. After that, the configured server is available to the client as usual.

What about Cursor, Codex, or another MCP client? They connect to https://api.plori.ai/mcp with ordinary OAuth 2.1 or an API key. The pairing-code flow on this page is Claude Code specific. Start from the connect guide.

Related

• Connect over MCP, for every other client and for API keys
• Getting started, for the first real run
• What a credit buys, for what does and does not bill