ARCY MCP
Connect your coding agent, Claude or ChatGPT once for the whole workspace, run one prompt that sets a Product up or battle tests your chat, and decide what it may read.
Your coding agent sets your Products up from your repository. You connect it on Settings > Install > ARCY MCP, once for the workspace. Claude and ChatGPT connect the same way and read your ARCY data.
Before you start
- Only an Admin presses Allow when an agent signs in, and only an Admin revokes one.
- One sign-in reaches every Product in the workspace. With one Product, nothing names it; with several, your agent asks which, then names it on every call.
- Point your agent at a local or staging copy: a preview calls your API for real.
- An agent reads and proposes; it cannot publish or delete, and the only thing it accepts is a Procedure, which stays off until you turn it on.
Setting it up
Connect your agent. Pick your client: four coding agents, Claude Desktop and claude.ai, or ChatGPT. No key goes in any command or file. Or press your agent's logo to copy a task that has it add the ARCY MCP. The two connectors take the address alone: a task is pasted into an agent working in a repository.
Allow the sign-in. Allow only a sign-in you started; the card marks an unverified client and names how many Products the agent reaches.
Connected agents. Every agent signed in, one token per client per person; signing in again replaces it and keeps its proposals.
Let a connected agent read visitor behaviour. Off until an Admin turns it on, which records who and when. While it is off your agent reads your setup and never what visitors did. What it may read once on is in What ARCY collects.
What your agent can do. One row per Product opens its entries. Set up ARCY end to end, marked Start here, covers one Product per run, then names the rest. It updates an old arcy.js with the Changelog's code changes, then runs all fourteen surfaces, each also its own entry.
What your agent can propose. One row per tool that writes a proposal, propose_procedures and accept_procedure among them.
What your agent can read. get_flagged_conversations returns last night's flagged Conversations with their transcripts, for your own routine.
Battle test the chat and fix what it finds. Your agent writes conversations a visitor of your Product would have, ARCY plays each against your Development chat, and your agent proposes only the fixes a re-run proved. A fix it proposes is accepted on Persona. It calls your public operations for real, which How AI request content is handled covers; it writes nothing.
Hand it to your agent. Every Hand it to your agent, ticked cards included, opens a menu: Open it in ARCY MCP lands on that card's entry here, and one entry per coding agent copies the config, the sign-in step and the line that runs that card's prompt.
Ask your developer. Beside that menu, for anybody who is not the developer. It gives a link to paste wherever your team talks to its engineers: the page behind it says what is needed, why, which page it lands on and how long the work takes. It needs no ARCY account, so a contractor can act on it, and holds no keys and no customer data. It expires after 30 days.
What it changes
Every write is a proposal, on the Product the call named, and the page that owns it is where Accept applies it with its own checks. Imported rows wait as an Import. A proposed Procedure waits in Proposed by your agent on Procedures, where a change to one you have lists what it rewrites. A Domains proposal is the workspace's and waits, one at a time, on Domains. No tool proposes the spend caps: those are set on Limits. Revoke discards open proposals, naming each Product; accepted work is kept.
Four prompts edit your repository, none touching git. Install arcy.js wires the Quickstart code and owns ARCY_TOKEN, ARCY_SECRET, ARCY_DEV_TOKEN and ARCY_DEV_SECRET, never a Secret key's value; with Update arcy.js, it gives an npm install its Update bot. Set up your events writes your arcy.track() calls against Events. Update your policies writes the Legal templates into your own. Remove ARCY takes it all out and clears returning visitors' storage.
Every tool but four takes a Product, left out when you have one: the workspace reads and the Domains proposals name none. One agent makes 180 calls a minute at most. Not on the MCP: file uploads, Require identity verification, DNS records, deleting a Product, and Accept on anything but a Procedure.
When it goes wrong
- The agent gets 401 or says it needs login. Sign in again from the agent (in Claude Code,
/mcp, then Authenticate). - A call is refused for naming no Product. Tell your agent which slug. If its tools offer no Product to name, it holds an old tool list: reconnect it. A slug refused as not found is misspelled, or its Product is Not serving; the dashboard URL carries the real one.
- The card says the link has expired. Ten minutes passed, or another tab answered it: start again from your agent.
- The card has no Allow. You are a Member: an Admin signs in from their own agent.
See also
Identity verification
Turn the identity check on once your signed identify() calls verify, and read the chart that shows how each session's claim fared.
Product
Set the Product's name, public name, URL, slug and icon, let the nightly check write on Development, pause ARCY in this Product, or delete it.