Connect an AI agent

Quick start

1. The server URL:

   https://www.promptdock.ai/mcp

2. In a terminal, add it to Claude Code:

   claude mcp add -s user --transport http --client-id a1b7de6e-807d-4374-b533-7b9163f357a9 --callback-port 8765 promptdock https://www.promptdock.ai/mcp

3. Then /mcp → Authenticate, approve in the browser, and paste the sentence under step 5.

Before you start

You need a Prompt Dock account; sign up first. Signing up adds an email confirmation, so finish it before you connect an agent: the agent waits on the browser while you do.

Prompts and skills submitted by members are reviewed by a person before they go live.

The server URL

One URL for every client. It speaks Streamable HTTP and signs you in with OAuth in your browser.

Claude Code

Checked with Claude Code 2.1.179.

1. 1. Add the server

   claude mcp add -s user --transport http --client-id a1b7de6e-807d-4374-b533-7b9163f357a9 --callback-port 8765 promptdock https://www.promptdock.ai/mcp

   -s user adds it for every folder. Without it Claude Code adds it to the current folder only.

   If you prefer the client to register itself

   claude mcp add -s user --transport http promptdock https://www.promptdock.ai/mcp

   With this command the approval page labels the app Unverified app: it registered itself, so Prompt Dock can't confirm who made it. That is expected for the real client too. Check that the name and the computer are yours before you approve.

2. 2. Authenticate

   In Claude Code, run /mcp, choose promptdock, then Authenticate. Your browser opens Prompt Dock's approval page; sign in if it asks.

3. 3. Choose what it can do

   Pick Create, edit and submit if you want it to submit for review; you can change this later in Settings → Account → Connected apps. Then approve.

4. 4. Confirm

   Back in Claude Code, run /mcp: the server shows Connected.

5. 5. Make the first call

   Paste this into Claude Code:

   Create a draft prompt titled "MCP connection test" with the body "Say hello to {{name}}." and give me its dashboard link.

   Expect a reply with a link to your dashboard, where a draft named MCP connection test now sits. If anything fails, ask it to run whoami: the answer names the account and the level it connected with.

Codex

Checked with Codex 0.161.0.

1. 1. Add the server

   codex mcp add promptdock --url https://www.promptdock.ai/mcp --oauth-client-id 3cb8ebcd-7bcc-4752-a1a1-4ce99d0763cc

   This client signs in on port 8766, which Codex reads from ~/.codex/config.toml. Add the port line at the top of the file:

   ~/.codex/config.toml

   # ~/.codex/config.toml
   # A top-level key: keep it above every [table] in the file.
   mcp_oauth_callback_port = 8766
   
   [mcp_servers.promptdock]
   # Written by `codex mcp add`; leave its lines as they are.
   url = "https://www.promptdock.ai/mcp"

   If you prefer the client to register itself

   codex mcp add promptdock --url https://www.promptdock.ai/mcp

   This form needs no port line.

   With this command the approval page labels the app Unverified app: it registered itself, so Prompt Dock can't confirm who made it. That is expected for the real client too. Check that the name and the computer are yours before you approve.

2. 2. Authenticate

   codex mcp login promptdock

   Your browser opens Prompt Dock's approval page; sign in if it asks.

3. 3. Choose what it can do

   Pick Create, edit and submit if you want it to submit for review; you can change this later in Settings → Account → Connected apps. Then approve.

4. 4. Confirm

   Start Codex and run /mcp: promptdock is listed with its tools.

5. 5. Make the first call

   Paste the same sentence as Claude Code's step 5, and expect the same reply: a link to a draft in your dashboard.

Other clients

Use a client that supports Streamable HTTP and OAuth. Desktop clients send no Origin header and work; hosted browser clients need an allowed origin. Configuration fields vary by client.

{
  "mcpServers": {
    "promptdock": {
      "type": "http",
      "url": "https://www.promptdock.ai/mcp"
    }
  }
}

Claude on the web and the Claude desktop app add it as a custom connector: Settings → Connectors → Add custom connector. Paste https://www.promptdock.ai/mcp as the server URL and name it Prompt Dock. Under Advanced settings, paste aa8c5f93-134d-4713-b304-518dda52d3c1 as the OAuth Client ID and leave the OAuth Client Secret blank.

Today only an MCP client can sign in and get a token: there is no device flow and no promptdock token command yet, so a script cannot call the server on its own.

Sign in and re-authenticate

Signing in opens Prompt Dock's approval page in your browser. That page is where you choose what the app can do:

• Create and edit (preselected): create and edit your prompts, skills and blog posts as drafts or unlisted, read everything of yours, and reveal public prompts and install skills under your plan's limits.
• Create, edit and submit: all of that, and submit prompts and skills for review. It is the level you need to submit.

Both levels can do things you may not expect: an unlisted item is reachable by its link; changing a published prompt's access, visibility or output applies at once; an uploaded image is public by URL the moment it lands; and a premium reveal or a completed install counts exactly as it would in the dashboard.

The app works only through Prompt Dock's own server: its token can't write to the database or to storage directly, so the level you chose is the whole of what it can do.

Each time you add the server the app registers again and asks again; earlier connections stay listed until you revoke them. Connecting an app signs it in to your account and we email you each time, so if you don't recognise one, revoke it in Settings → Account → Connected apps: it is signed out of Prompt Dock's server at once.

• If the browser shows "can't be reached" right after you approve, your agent stopped waiting: run the login command again and approve once more.
• A sign-in error that appears before Prompt Dock's page loads comes from Supabase's authorization page, which runs our sign-in. It carries no Prompt Dock reference; run the login command again to retry.
• If the approval page says the connection request expired (it happens when you signed up mid-way and confirmed your email first), go back to your agent and run the login step again.
• If you add an authenticator to your account after connecting an app, its calls are refused with mfa_required: revoke the app and connect again, and you'll be asked for your code.
• When a tool says you are signed out, sign in again: in Claude Code run /mcp and choose Re-authenticate; in Codex run codex mcp login promptdock.

What the tools do

Every argument, result field and an example call per tool are in the tool reference.

Limits

Submissions are limited to 5 an hour and 20 in any 24 hours because a person reviews each one. Submit again after the wait, or from the dashboard. Both windows roll: each submission counts for an hour and for 24 hours from the moment it was made. Past the budget, open the draft in your dashboard and submit it there; a submission refused for the budget changes nothing, and the drafts you already made are saved.

New prompts and prompt reveals are counted per member over a rolling hour. A call refused before it does anything takes nothing: a create with a bad field, or a reveal of a prompt that isn't there or is above your plan. A create with publish: true counts as a new prompt and as a submission. When a limit is reached, the error says how many seconds until the next call goes through (retry_after_secs); the dashboard is never limited by your apps.

Every other tool has its own budget, kept apart from your dashboard's, so an agent that loops can spend only its own.

Large uploads

A call can carry 3 MB of decoded inline bytes, and the hosting platform refuses a request body past about 4.5 MB before it reaches us (the client sees a bare HTTP 413). Anything from disk goes through begin_upload instead: it returns an upload link and a ready-to-run curl line per file, then the create names the uploaded file. An image is public by URL as soon as a create or update saves it to the item, before any review. Removing it stops the item showing it, but its URL can keep loading afterwards, so never put anything private in an image.

You need a longer client timeout only when a call carries a skill package or a set of images large enough to keep the server working past 60 seconds; every other tool answers well inside the default.

[mcp_servers.promptdock]
tool_timeout_sec = 180

Claude Code: start it with a longer MCP tool timeout (in milliseconds).

Tell your agent

Paste these five lines into your project's CLAUDE.md or AGENTS.md:

If something goes wrong

The two error shapes

A tool that can't do what was asked returns a result marked isError: first a sentence with what went wrong and what to do, then a JSON object with code, message, fix and doc_url. Before any tool runs, the server can also answer over HTTP: a 401 makes the client ask you to sign in again, and a 403, 429 or 503 carries the same JSON object. Every code is explained in MCP errors.

A fault on our side carries a reference that starts with mcp_. Send it to support with roughly when it happened; it finds the server log line with the cause. To disconnect an app, revoke it in Settings → Account → Connected apps.