AI assistants & MCP — Help center

What the connection does

Phloz ships a built-in MCP server (Model Context Protocol — the open standard AI assistants use to talk to software). Connect it once and the assistant can read and act on your workspace from inside its own chat: clients, tasks, messages, each client's tracking-infrastructure map and Google Ads, and help decide what to do first.

23 tools, 9 ready-made prompts and 6 read-only data views. Seventeen tools read: clients, stale clients, one client's profile, tasks, one task in full, the team, messages, a tracking map, its health summary, a full tracking audit, and the decision reads below. Six write: create a task, change a task's status, comment on a task, add or update a tracking node, draft a reply to a client.
Decision reads. A portfolio briefing triages every client you can see (urgent, soon, watch) with the numbers behind each flag; a client brief puts one client's open work, waiting messages, retainer hours, tracking audit and Google Ads in one answer. On plans with Ads Health, the assistant can also read the Google Ads alerts from the morning check, an account's daily figures and campaigns, the weekly audit with the lists behind it (search terms, negative-keyword candidates, keywords, campaigns), and month-to-date spend against the budget agreed with the client; see how those work. These are facts with their numbers, not verdicts: the order is a stated rule, and you decide.
Scoped like you are. The assistant acts with your role and your client assignments, never more. It sees one workspace per connection and nothing from any other agency's workspace.
Writes are conservative. Tasks it creates are internal-only, never client-visible. A drafted reply lands in Messages → Drafts for a person to review and send. The one write that reaches clients is a stage change on an item they see (see permissions). Every request it makes is recorded in the Activity log under your name and the connection's.
Available on the Growth plan and above — see pricing.

Connect from claude.ai

No token needed. Claude registers itself and walks you through signing in and approving one workspace.

In claude.ai open Settings → Connectors and choose Add custom connector.
Paste the server URL https://app.phloz.com/api/mcp. Leave OAuth Client ID and Client Secret empty.
Continue. You are sent to Phloz to sign in (if you are not already), then to an approval screen showing the app's web address. An app that registered itself, as claude.ai does, is marked Unverified app; that's expected when you have just started the connection yourself.
Pick the workspace and the access level: Read-only (view clients, tasks, tracking maps and messages you can access) or Read & write (also create tasks, edit tracking maps and draft replies, bounded by your role and assignments). Click Allow access.
Back in claude.ai the Phloz tools appear in the chat. In Phloz the connection shows under Settings → AI access (MCP) as “…(OAuth)” for as long as it can renew, used or idle, and you can revoke it from that row at any time.

A read-only connection is offered only the seventeen read tools, so the assistant is never shown a write it cannot make. To add writes later, reconnect and choose Read & write.

Claude Code, the Claude API, ChatGPT, Grok, Gemini and other MCP clients

Any client that supports remote MCP servers over HTTPS with the streamable HTTP transport can connect to the same URL (clients that only speak the older SSE transport can't). Two paths:

OAuth: if the client supports it (claude.ai does; check your client's documentation for ChatGPT, Cursor, Grok or Gemini), add https://app.phloz.com/api/mcp as a custom MCP server and follow the same sign-in and approval flow as above.
Personal access token: for Claude Code, the Claude API's MCP connector, or any client that takes a bearer credential. In Phloz open Settings → AI access (MCP), create a token (read-only or read & write; it never expires, or lasts 30 days, 90 days or a year), copy it once, and send it as Authorization: Bearer phloz_mcp_… to https://app.phloz.com/api/mcp. The token is not an OAuth client secret; do not paste it into a Client ID or Client Secret field.

Tokens act with your role and client assignments, and are revoked from the same settings card. Give each token a name you'll recognise; a token reaches every client you can, and can't be limited to one client.

Things to ask it

Plain language works; the assistant picks the tool. A few that exercise different parts of the workspace:

“Which of my clients have had no activity in the last 30 days?” — lists stale clients, most neglected first.
“Run the tracking audit on Thornbeck Outdoor and tell me what is broken or unverified.” — the same checks the app runs, with a score and the findings that need attention. Rules you've snoozed on the client's Audit tab are left out, as in the app.
“Create a task for the PPC department on Driftmoor Dental: rebuild the purchase conversion action, due Friday.” — a new internal task, attached to the client (read & write connections only).
“Show me the last three messages from Driftmoor Dental and draft a reply confirming the GA4 fix is live.” — reads the thread, then stores a draft in Messages → Drafts for you to review and send.
“What is on the SEO department's plate this week, and what is overdue?” — the department's tasks with their due dates and which are overdue. The assistant works out “this week” from your workspace's own date.
“Which clients need us most this week, and why?” — the portfolio briefing, then a brief on each urgent client, with the facts behind every point.
“Why did Driftmoor Dental's Google Ads leads drop?” — the account's daily figures, its alerts, the weekly audit and the tracking audit, with the likely causes ranked (plans with Ads Health).
“Which search terms should we negate for Driftmoor Dental?” — the weekly audit's wasted terms as a paste-ready negative list per campaign, after checking conversion tracking can be trusted. You apply it in Google Ads; Phloz never changes the account.
“Who is off pace on ad spend this month?” — month-to-date spend against the monthly budget on each client in Phloz, with the projected month end.
“Answer what came in from Thornbeck since yesterday.” — reads the new messages and any request they mention, then saves a reply draft on each conversation for you to send.

Nine ready-made prompts ship with the server too: a weekly portfolio review, diagnose a client's Google Ads, find negative keywords, check ad spend pacing, answer the client inbox (drafts for you to send), prepare a client update (a draft), audit a client's tracking, plan a department's week, review stale clients. Clients that support ready-made prompts list them alongside the tools.

Permissions and limits

Never more than you. Owners and admins can use every tool across the workspace; members read and write on the clients assigned to them; viewers read assigned clients and can comment on tasks. If your role changes, the connection follows your new role at once; if you're removed from the workspace, it stops working.
Read-only means read-only. A read-only connection or token cannot create, change or draft anything.
What it can never do: write or send a message to a client (a reply is a draft a person sends), change Google Ads, change billing, invite or remove people, delete a client, or see another workspace.
One write reaches clients, as it does in the app. Moving a request or a shared task to a new stage sends its contacts the same progress email as moving it in Phloz, about two minutes later; contacts who turned progress emails off are skipped. The tool tells the assistant so before it moves anything, so it can ask you first, and its answer says whether an email goes.
Everything is on the record. The Activity log lists every request the assistant makes, reads included, as “your name via AI assistant”, with the connection, the tool, whether it went through and, for a write, what it wrote (a task's title, a node's label, a comment). Anything it creates shows you as the author.
Each connection has a request limit, so one runaway assistant can't exhaust the workspace.
Keep a client away from AI assistants. On the client's page, an owner or admin can switch on “Hidden from AI assistants”: connected assistants then can't see the client, its tasks, messages, tracking map or Google Ads, and lists only say how many clients are hidden. Your team still sees everything in Phloz.
Pause AI changes. In Settings → AI access, an owner or admin can pause every change from AI assistants at once: creating or moving tasks, comments, map edits and reply drafts are refused (and logged as refused) until an owner or admin turns the pause off. Reading goes on.

How the connection handles your data

What Phloz stores about a connection: the connection or token name, its access level, when it was created and last used, and a scrambled fingerprint of the key, not the key itself (a token is shown once and never stored). Every request adds an Activity log entry.
What leaves Phloz: only the answers to the calls your assistant makes, sent to the AI provider you connected — claude.ai, OpenAI, xAI or another client. How they retain and use what they receive is governed by their terms, not ours. Phloz never sends workspace data to an AI provider on its own.
Lifetimes: a connected assistant's access renews itself every hour. A connection expires after 30 days without use, and every connection needs approving again 90 days after you first allowed it. Expired keys are deleted daily. Personal access tokens last until you revoke them, or until the expiry you picked when you created them.
Connections are always encrypted, and requests from web pages other than Phloz and the AI apps we know are refused before any key is checked.
The rest of our posture — workspace isolation, export, deletion — is on Security & data, with the details in the privacy policy and the DPA.

Troubleshooting

“Requires Growth plan or above” on the approval screen: the workspace is on a plan without AI access. Upgrade under Billing, then connect again.
Unauthorized: the token or connection was revoked or has expired, you were removed from the workspace, or the workspace moved to a plan below Growth. Reconnect from the client, or create a new token.
The write tools are missing: the connection is read-only (reconnect and choose Read & write), or your role is Viewer, which can only comment on tasks.
“not_found” for a client you can see in the app (the assistant says it can't find the client): the token acts with the access of the person who created it, and that client isn't one of theirs, or an owner or admin switched the client to “Hidden from AI assistants” on its page (the clients list marks it “Hidden from AI”).
Anything else: contact support with the connection name and the time of the failed call.