Use Copilot from Claude, Cursor and other AI apps
Flowcall Copilot is available as a remote MCP server, so you can work with it from an AI app that supports MCP connectors, such as Claude, ChatGPT, Cursor or VS Code. The app talks to Copilot for you: it starts Copilot conversations, answers Copilot's questions, checks the saved configuration and, when you ask, saves and publishes task changes.
The app acts as you. It can only reach the accounts you can reach and do what your role allows, and the conversations it starts appear in your Copilot history in the dashboard.
Connect an AI app to Flowcall Copilot
You need Copilot enabled for your user. Then:
-
In your AI app, add a custom MCP connector (it may be called a remote MCP server or custom connector) with this URL:
https://api.flowcall.co/mcp -
When the app asks you to sign in, it opens Flowcall in your browser. Sign in if needed.
-
On the Authorize Copilot access page, check that Client is the app you just connected and Returns to is that app's address (for example
claude.ai), then choose Authorize. The browser returns to the app.
The connection lasts 30 days. After that the app asks you to authorize again. An authorization page that is left open expires after 10 minutes; connect again from the app to get a new one.
Only authorize a request you started yourself. If you open an authorization page you didn't expect, choose Deny.
Ask Codex or Claude Code to connect Flowcall Copilot
Copy this prompt into Codex, Claude Code or another coding assistant that can configure MCP connections:
Connect this AI app to Flowcall Copilot using the remote MCP server:
https://api.flowcall.co/mcp
Use this app's supported Streamable HTTP MCP setup and OAuth sign-in.
Reuse an existing Flowcall connection if one is already configured.
Start sign-in and let me approve access in the Flowcall browser page.
After authorization, load the Flowcall tools and list the accounts I can access
to verify the connection. If I need to reload the app or open a new chat first,
tell me how. If setup fails, show me the exact error.
Setup guide: https://docs.flowcall.co/copilot/mcp/
You still need to sign in and choose Authorize yourself. If the assistant cannot configure MCP connections, add the server URL in your app's MCP settings using the steps above. A prompt can help configure the app, but a connection error may need a fix from Flowcall or the app provider.
What an AI app can do with Copilot
Through the connector, the app can:
- list the accounts you can use and start or continue Copilot conversations on them;
- send your requests to Copilot, answer its questions and read its replies, plan and proposals;
- read the saved configuration to check Copilot's work: tasks, FAQs, workflows, data points, sample data, Custom APIs, and whether saved tasks match what's live;
- read one live task with each objective labelled by where its value comes from (the customer's answer, inferred by the AI, a data point, a computed value, an action, an attachment or an agent) and whether the AI reads it from the conversation. See What the AI reads on each turn;
- search the Flowcall docs, so the app explains product behavior from the docs instead of guessing;
- accept or reject Copilot's task proposals, and publish a saved task to its live flow;
- undo a change Copilot applied in a conversation (see Undo a Copilot change), when you ask it to;
- for account admins, read debugging evidence for a reported issue or a customer message.
- for account admins, look up tickets, customers, messages, Custom API calls and template sends, count tickets for a period, and list the tickets of a period for a report. See Look up account data from an AI app.
- read the Tickets page Insights: breakdowns and pivots with rates such as % resolved by AI or median time to resolve, and a task's funnel. See Read ticket insights from an AI app.
- read Conversation Audit findings and Automation Gaps, with counts by issue type and a link to each conversation. See Read audit findings and automation gaps from an AI app.
- read why customers are unhappy: each unhappy ticket's main reason, who can act on it and how it ended, with the tickets behind them. See Read why customers are unhappy from an AI app.
Other proposals, such as data points, Custom APIs, FAQs, templates and tests, are reviewed in the dashboard. The app gives you a link that opens the same conversation in Copilot.
Long Copilot requests keep running in Flowcall while the app waits. The app checks back for the reply instead of sending the request again, so a slow request is never applied twice. If you type or save something in the same conversation in the dashboard, the app reads those new messages before it sends anything else.
Look up account data from an AI app
Account admins can ask the app about real records, for example "why didn't ticket FC-1234 get a refund?" or "what did this customer get on WhatsApp yesterday?". The app can read:
- A ticket, by its ID or exact name (such as
FC-1234, or just1234): status, task, team, assignee, child and interaction tickets, the latest 50 history entries, notes and collected values. - A customer, by customer ID, phone number (any format) or email, with their 20 latest tickets.
- A customer's messages on WhatsApp, the chat widget, Instagram and email, with delivery status and errors.
- Custom API calls and actions for a ticket or customer: what was sent, what came back, and any error.
- Template sends and events for a phone number or customer: which WhatsApp templates were triggered, whether they were sent, delivered or failed, and the incoming events (such as order updates) behind them.
- Ticket counts for a period, for questions like "how many tickets did each task get last month?" or "how many Estimated_delivery_date tickets had the order already delivered?". See Count tickets from an AI app.
- Ticket lists for a period, for reports like "this month's reopened tickets" or "tickets with more than 4 customer interactions". See List tickets from an AI app.
To keep your account fast for agents and customers, these lookups are narrow:
- Each record lookup starts from an ID, phone number, email or exact ticket name. There is no free-text search or name search.
- Lists cover a date window of up to 31 days (7 by default) and return at most 50 rows. Long messages and API payloads are shortened.
- Each lookup has a few seconds to finish, and each person can run 30 lookups a minute.
Count tickets from an AI app
Account admins can ask the app for exact ticket totals instead of estimates from a handful of tickets. A count covers tickets created in a period, leaves out test tickets, and can be:
- grouped by task, status, channel, day, week, or the values collected for one objective;
- filtered by task names, statuses, channels, customer conversations only (root tickets) or child and interaction tickets only, and by the values collected for one objective, such as tickets of
Estimated_delivery_datewherestatus_of_orderisDelivered.
Limits:
- One count covers at most 92 days. Give dates (in your account's time zone) or a number of recent days; the default is the last 30 days. Split longer periods into several counts.
- Counting by an objective's values needs exactly one task, and only counts tickets where that objective has a stored value. Objectives that come from an Objective Collection can't be counted by value yet.
- Grouped results show the 100 largest groups; the total still covers all of them.
- A count can take up to 20 seconds and runs one at a time. The same question asked again within 15 minutes returns the saved answer.
AI usage and costs are not available through the app.
List tickets from an AI app
Account admins can ask the app for the tickets themselves, for example to build a sheet of this month's reopened tickets, tickets with a negative sentiment, or tickets where the customer got in touch many times. A ticket list covers tickets created in a period, newest first, and leaves out test tickets. Each ticket comes with:
- its name, task, status, channel, created and resolved time, assignee, tags and a link to the conversation;
- its reopen count and sentiments;
- for customer conversations (root tickets), the same counts as the Chat Interactions, Customer Email Messages and Incoming Calls columns of the ticket export, across the ticket and its child and interaction tickets, plus their total.
The list can be filtered by task, status, channel and ticket type (customer conversations only, the default, or child and interaction tickets only), and by:
- a minimum reopen count, such as reopened at least once;
- sentiments, either specific ones or any sentiment except the ones you name, such as any sentiment except AI Disabled;
- a minimum number of customer interactions (chat interactions, customer email messages and incoming calls added up).
The app can ask for tickets that meet all of these filters, or any of them, which gives a single list of unique tickets across several conditions.
To protect your customers' data, a ticket list never includes the customer's name, phone number or email. The app can look up one ticket or customer when you need those details.
Limits:
- One list covers at most 92 days. Give dates in your account's time zone, or a number of recent days.
- The app gets up to 100 tickets at a time and asks for the next part until it has them all. Each part counts toward the limit of 30 lookups a minute, so long lists take a little while.
- When the list is filtered by customer interactions, each part checks up to 1,000 tickets, so a part can come back with only a few tickets while more remain. The app keeps going until the list is complete. The total number of matching tickets is shown up front, except when the list is filtered by customer interactions.
Read ticket insights from an AI app
The app can read the same numbers as the Insights strip on the Tickets page, for questions like "which tasks have the lowest share resolved by AI?", "how long does each team take to resolve tickets?" or "where do NPS conversations go, and how many stop?". It counts the tickets you can see on the Tickets page, so anyone with Copilot access can ask, not only account admins.
- Breakdown: tickets grouped by status, channel, team, task, mood, child ticket objective or an objective's answers, optionally split by a second field, as a table of counts or of one measure: % resolved, % resolved by AI, % with an agent, voice FCR, median time to resolve or median first agent response. The app gets the counts each rate is taken over. It can ask for up to three breakdowns at once.
- Task funnel: the customer conversations that went through a task and the tasks they were routed to next, with how many went further, how many had an agent involved, and how often each objective was answered (for yes/no objectives, how often the answer was yes). See Task Funnel.
Both can be filtered by period, task, team, status, channel, ticket type, mood, objective answers, AI only or with an agent, open, unassigned and reopened tickets. Each answer comes with a link that opens the same view on the Tickets page. The page filters by one task and one channel, so when you ask about several, the link shows only the first.
Limits:
- One answer covers at most 92 days; the default is the last 30. Give dates in your account's time zone or a number of recent days.
- Each breakdown lists its 12 largest groups and says when there are more.
- Insights answers run one at a time with the ticket counts, so a busy moment can take a little longer. The same question asked again within 15 minutes returns the saved answer.
- For trends by day or week, the app uses ticket counts instead.
Read audit findings and automation gaps from an AI app
Teammates who can open Conversation Audits or Automation Gaps in Insights can ask the app which problems come up most, for example "what were the most common audit issues last month?" or "which customer questions does the AI keep failing to answer?". The app can read:
- Audit counts: how many audited conversations had findings in the last 1 to 180 days, split by analysis group, tag, stream, severity, trigger and channel. Findings are counted in the period they were first detected.
- Audit findings: the conversations behind those counts, filtered by group, tag, trigger, severity, stream or channel. Each finding comes with why it happened, quotes from the conversation and the suggested next action.
- Automation gap counts: gaps by type (missing task, missing FAQ, improve an existing task or FAQ, business description) and review status. Repeats of the same customer message count as one gap; the app also gets the number of messages behind them.
- Automation gaps: the customer message, what the AI concluded, the suggested fix and the closest existing tasks and FAQs, most repeated first by default. Gaps are kept for 90 days.
Every finding and gap comes with a link that opens the conversation in the dashboard, and a link to the finding or gap itself. The links switch the dashboard to the right account first.
The app can only read audits and gaps. Feedback on findings, review status changes and audit settings stay in the dashboard.
Read why customers are unhappy from an AI app
When Unhappy Customer Reasons is turned on for the account, teammates who can open CSAT in Insights can ask the app why customers are unhappy, for example "why is our CSAT low this month?" or "why do customers threaten to escalate?". These are the reasons shown under Why customers are unhappy in Conversation Audits → Unhappy customers. A ticket is unhappy when it has a low rating or the customer was frustrated, aggressive, threatened to escalate or disputed an outcome. The app can read:
- Unhappy customer reasons: for the last 1 to 180 days, unhappy tickets by signal and how many have a reason, split by main reason, by who can act on it (support setup, agent team, business operations), by how it ended (recovered or not) and by task. It also gets, for all ratings, how often they are low for AI-handled and agent-handled tickets, by the wait for the first agent reply and by the time to resolve.
- Unhappy tickets: the tickets behind those counts, newest first, filtered by signal, reason, owner, ending, task or AI vs agent. Each comes with its signals, the customer's rating comment, a one-line summary, quotes, the audit findings behind the reason, measured waits and a link to the conversation.
Support-setup reasons can be turned into changes through Copilot conversations. The app can only read the reasons; it cannot change them.
Saving and publishing changes from an AI app
The app only saves or publishes when you ask it to. Flowcall applies the same checks as the dashboard:
- Simple changes (wording, triggers, flags, dispositions) are saved directly.
- Logic changes (objectives, conditions, transitions, possible values) save directly only when the task's linked tests pass. Otherwise Flowcall asks the app to get your explicit confirmation first, and the app should show you the reasons.
- Validation errors are never skipped. Nothing is saved, and the app sends the problems back to Copilot to fix.
Accepting a task proposal saves it to your business context. It goes live only when it's published. Data points and Custom APIs work differently: they take effect for linked tasks as soon as they're saved in the dashboard.
Never paste API keys, tokens or passwords into the AI app. Enter them in the integration's settings in the dashboard.
Troubleshooting the Copilot connector
- "Needs Auth" or "Authenticate": the server is configured, but sign-in has not finished. Choose Authenticate, sign in to Flowcall and authorize access. If authentication shows an error, use that error to identify the problem.
- "Dynamic Client Registration rejected" or "Only the authorization_code grant is supported": the app's automatic registration failed before sign-in. Retry authentication. If it still fails, contact Flowcall support with the exact error and your app's name and version.
- "AI Copilot is not enabled for your user": ask an admin on your account to turn on Copilot for you, then try again.
- An account is missing from the list: the app only sees accounts you can open in the dashboard. Check that you can switch to it there.
- The authorization page says the code is invalid or expired: the request was already used, denied, or left open for more than 10 minutes. Remove and add the connector again, or use the app's reconnect option.
- The app stops working after a month, or after you revoked it: connections last 30 days. Reconnect and authorize again.
- A data lookup says it took too long or too many are running: ask for a shorter date window, or start from a ticket or customer ID. For a ticket count or ticket list, use a shorter period, or wait a few seconds for a running count or list to finish. If you hit the limit of 30 lookups a minute, wait a minute and try again. Lookups are for account admins only.
- Saving or publishing is refused: the app shows why. Validation problems go back to Copilot to fix; logic changes without passing tests need your explicit go-ahead. See Saving and publishing changes from an AI app.
Revoke an AI app's access
Open https://dashboard.flowcall.co/personal-api/authorize while signed in to Flowcall. Your Copilot credentials lists every connected app and Copilot helper with when it was last used. Choose Revoke to disconnect one immediately.
Copilot MCP connection details for developers
These details are for building or configuring an MCP client yourself. Most AI apps only need the URL.
- MCP endpoint:
https://api.flowcall.co/mcp, Streamable HTTP. Requests arePOST; the server does not offer a separate event stream or sessions. - Sign-in: OAuth 2.1 authorization code flow with PKCE (
S256only), for public clients without a client secret. Clients register themselves through dynamic client registration. - Registration grant types: clients may request
authorization_codewith optionalrefresh_tokenmetadata. The registration response grants onlyauthorization_code; Flowcall does not issue refresh tokens. - Desktop callbacks: HTTP callbacks to
127.0.0.1or[::1]may choose a different local port at sign-in. The host, path and query must still match the registered callback; other redirect URIs must match exactly. - Discovery: an unauthenticated request to the endpoint returns
401with aWWW-Authenticateheader that points tohttps://api.flowcall.co/.well-known/oauth-protected-resource/mcp. Authorization server metadata is athttps://api.flowcall.co/.well-known/oauth-authorization-server. - Scope:
copilot. Access tokens last 30 days and there are no refresh tokens, so clients authorize again when a token expires. - Who the token acts as: the teammate who authorized it, with their current permissions. Account access tokens from Settings → Integrations don't work with the Copilot connector.