Skip to main content
You can use Connie from the AI app you already work in, such as Claude, ChatGPT, Claude Code, Cursor or VS Code. Once connected, you can ask it to work with data in your Connie workspace. ChatGPT and other MCP clients can discover and run Connie’s public tools, including Workflow edits, with your workspace permissions. Some tools require a Connie browser tab or an originating Connie request.
Before opening Connie Settings or copying an API key: finish the first run of a new workspace. If Settings or a direct ?tab=developers link returns you to it, answer its two questions and try again. New workspaces use free credits, with no card or trial. Plugins remains available during setup; Connect my agent and Settings are not setup exceptions. See Settings access during setup and API keys.
This works through MCP. Connie runs an MCP server at https://mcp.connie.ai. Follow your app’s guide to install Connie from its directory or add the server address, then sign in to Connie. ChatGPT’s plugin and Claude’s directory connector handle the server address for you. Where: the agent menu (click the agent card at the top of the sidebar) → Connect my agent, or Settings → Developers → Connect my agent. After setup, the direct link https://connie.ai/<workspace-id>/connect-agent also works.
What is MCP? MCP (Model Context Protocol) is an open standard that lets AI apps plug into other services. A phone charger is a good comparison: any phone with the right port can use any charger with the same plug. In the same way, any AI app that supports MCP can connect to Connie. Each app calls it something different: a connector in Claude, an app or plugin in ChatGPT, an MCP server in Cursor and VS Code.
Connect stays available after setup. If you pick your AI app during setup (Which agent do you use?), Connie stays in front and opens the app’s setup behind it: Claude’s directory listing, ChatGPT’s plugin listing, or the guide for another app. If the new tab is blocked, the guide opens instead. When a task is supplied with the setup link, Your first task keeps it ready to copy and send in your AI after connecting. The sidebar has no Meet the Connie MCP card any more: Connect my agent is in the agent menu. Claude is added from Claude’s own connector directory, so its card has two buttons: Connect opens Connie’s page in Claude’s directory in a new tab, where you click Connect to Claude, and Documentation (shown when you point at the card) opens the step-by-step guide in Connie. If your browser blocks the new tab, the guide opens instead. ChatGPT has one card for personal and workspace accounts: Connect opens Connie’s plugin page in a new tab, where you click Install plugin. Documentation opens the eight-step guide in Connie. It is available on hover, by keyboard and on touch screens. If the new tab is blocked, the guide opens instead. The Support button in the header opens our support chat. If support chat is unavailable, it opens the help site instead. On the standalone Connect page, use Connect to browse platforms and My accounts to see your existing connections and each account’s status. Both lists scroll while the title, tabs and search stay in place. Account cards use the same compact size as platform cards. Here and in Plugins → Connect my agent, while Connie checks whether an app is already connected, its card says Checking… instead of offering Connect; it then shows Manage, Reconnect or Connect as appropriate. When there are no accounts, the centered “No connections yet” empty state offers Connect an agent, which opens the Connect tab. If the account list shows “Could not load your connections.”, click Try again. The connection check remains available inside individual setup guides. Connect my agent in the agent menu opens the standalone Connect page. The standalone Connect page and the Settings page keep their Connect and My accounts tabs; their platform search box says Search platforms…. In My accounts, click Manage to open that connection’s controls in Plugins; the back arrow at the top left returns you to Connect my agent, on the tab you came from. Click Reconnect on an expired connection to open its setup guide; Codex and Cursor open the Manual tab for signing in again, and an unrecognized client opens Other MCP clients. Check connection checks your sign-in for the agent whose guide is open. A Slack connection or a teammate’s sign-in does not mark your Claude, ChatGPT or Codex sign-in as connected. For example, the Claude guide shows “Your Claude sign-in is connected.”, “Your Claude sign-in needs reconnecting.”, or “No Claude sign-in found for your account. This page updates automatically.” Slack’s guide checks the workspace’s Slack connection. Other MCP clients asks you to verify in your agent because Connie cannot identify every custom client’s name. Guides that offer API-key setup show “API-key connections are verified in your agent.” below the Manual sign-in status. The guide does not infer API-key connection status from OAuth sign-ins.

What you can do with Connie in an AI app

Ask in plain language, for example: “Find 10 VPs or Heads of Sales at US B2B SaaS companies with 50–500 employees, and go ahead and save them to a new list in Connie.” Your AI app can:
  • Records and Lists: search, read, create, update and delete People, Companies, Deals, Tasks and your custom objects, and manage lists. It can restore records deleted in the last 7 days. When it links records (a person’s Company), it can add a link and keep the others or remove just one; a plain list of records replaces the links. See Records.
  • CSV import and export: bring a CSV into a list or export one.
  • Find leads and enrich: search for people and companies, and add emails, phone numbers and company data. See Lead Finder and Enrichment.
  • Workflows: build, edit, test, run and turn on workflows. See Workflows.
  • Send email and messages from the accounts connected to your workspace.
  • Mute a LinkedIn conversation (when released in your workspace’s catalog): use inbox-search and inbox-get-conversation to identify one exact saved Classic direct/group conversation, then inbox-set-muted with its conversationId and muted: true or false. Clarify ambiguous targets/accounts. Use the selected account’s accountRef in the host envelope when required. Shared account access, policy and billing apply. No Workflow Actions/Agents, Records columns or program execution. Never automatically repeat an uncertain change or an accepted change awaiting Inbox refresh. See LinkedIn conversation controls.
  • Configure account sending limits: view and update daily caps, sending hours, time zone and warm-up settings for accessible LinkedIn accounts and email inboxes, or read every sending account’s capacity, state and daily target-versus-sent record in one call (leave the account out). Name the account and the settings to change. Limits use the account’s time zone (UTC when it has none) and remain separate from workflow running hours. See sending-limit settings.
  • Use your connected apps, such as your CRM or Slack. These actions happen outside Connie.
  • Manage your Skills: search, read, create, rename, edit, duplicate, disable or delete workspace skills; manage supporting files; download content; and enable or revoke public sharing. Built-in skills stay read-only. See skill management examples.
  • Search what was said in your meetings: ask “Why did Northwind push back on pricing?” or “What exactly did Jordan say about the renewal date?”. Your AI app searches the transcripts of your recorded meetings by topic or by exact words, gets back only the lines that match (speaker, time and the lines around them) with a link that opens each one in Connie at that moment, and can read on around a line instead of the whole transcript. See Ask across your meetings.
  • Read workspace details, such as members, files and billing.
In general MCP clients, ask for a meeting’s recording to get its status and a link to open it in Connie. calls-get-recording returns the meeting-page link in url, not a file download. Opening it requires browser sign-in and access to the meeting; the link grants no access. This tool is not available in the ChatGPT connector’s focused tool set. See meeting recordings. For Add record and single or bulk Object creation, pass the destination List UUID as listId. A missing List or an Object ID in place of a List is refused before Records are created or merged: “Choose a List before adding records. Nothing was created.” Find the intended List first, or ask which List to use when the destination is unclear.

Connect your AI app

Pick your app. Each uses your Connie sign-in. ChatGPT installs from Connie’s plugin listing; apps with manual setup use the server address https://mcp.connie.ai.

Claude

Web, desktop and mobile. Find Connie in Claude’s connector directory.

ChatGPT

Install Connie from ChatGPT’s plugin directory on web or desktop.

Claude Code

One command in your terminal.

Cursor

One-click install link, then sign in.

VS Code

For GitHub Copilot agent mode.

Other apps

Codex, Gemini CLI, Windsurf, Grok, Devin, Cline and any other MCP app.
Connie also has a step-by-step guide with screenshots for each app inside the product. Open Connect my agent (https://connie.ai/<workspace-id>/connect-agent) and choose your app. For saved-work verification, canvas-get, object-list-records and search-record accept fresh: true. In Connie chat this bypasses reuse of an earlier result from the request; direct MCP reads already query storage. The option does not change access, scope, filters, pagination or read-loop limits. Copy a newly created Page’s canvasId from its creation result to read it back; a read failure does not mean creation failed.

How connecting works

1

Add Connie to your AI app

Add Connie from the app’s directory, or add a custom connector. For ChatGPT, install the published Connie plugin; for other AI apps that need a server address, use https://mcp.connie.ai. You don’t need a client ID, a client secret or an API key.
2

Sign in to Connie

Your browser opens Connie. Sign in if you aren’t signed in already.
3

Choose a default workspace and allow access

The screen Access your Connie workspace via MCP opens. Pick a Default workspace and click Allow access.
4

Go back to your AI app

It now shows Connie as connected. Turn Connie on in a chat if your app asks you to, and start asking.
The access screen lists what the app will be allowed to do:
  • Read and send in your inbox: “View messages, calls, and threads — and send and reply on your behalf.”
  • Read and edit Records: “View and modify records in system and custom objects.”
  • Manage and run workflows: “Create, edit, and run automations on your behalf.”
This way of signing in is called OAuth. You approve access in your browser, and your AI app never sees your Connie password. You can take access back at any time from Connie. See Manage access.

Workspaces

Read the access screen when you connect. Where available, choose Only this workspace to restrict this authorization; All workspaces you belong to remains the initial selection. If it says only the selected workspace, the connection is limited to that workspace; a request or server URL cannot switch it to another one. If it says all workspaces you belong to, your sign-in covers every current workspace membership. For an all-workspace connection, the Default workspace you pick on the access screen is only where requests go when you don’t name a workspace. To work in another one, name it in your request, for example “…in the Acme Sales workspace”. See Manage access for switching and for changing the default.

Credits

Actions you run through MCP use credits from the workspace’s balance, exactly as they would inside Connie. A lead search through Claude costs the same as the same search in Connie.
  • Looking things up is free: searching Connie’s tools and skills, and reading a tool’s details or price.
  • Your AI app gets a receipt with each action that shows how many credits it charged, including any lookups the action ran itself (for example a Company Identity lookup inside Find People at Company), its AI usage and a Lead Finder import’s charge.
  • MCP actions show up in the credit usage log with the label MCP, under the person who connected the app.
  • The workspace’s balance is shared with everything else in Connie. If it dropped more than your receipts add up to, ask your AI app how many credits you have: owners and admins get the last 24 hours split by where the credits went, for example a Chat reply in Connie at the same time.
  • If the workspace runs out of credits, the action doesn’t run. Your AI app tells you what to do next. Owners and admins get a link to add credits or change plan. Members and guests are told to ask an owner or admin. When the workspace has used its free credits, owners and admins are told “Choose a plan to keep going (the free credits are used up)” and members “Ask a workspace owner or admin to choose a plan; the free credits are used up”. An Internal workspace receives a credit notice without a subscription or pricing link. Beta workspaces can use existing credits without a subscription; once those credits run out, an owner or admin can choose a plan.
  • If the workspace’s trial (started before November 2026) was cancelled, nothing that uses credits runs until it’s resumed. Your AI app says the trial is cancelled and gives owners and admins a link to resume it, with the days and credits it has left.
  • When fewer than 100 credits are left and auto top-up is off, your AI app gets a low-balance notice after a successful action.
Ask “What plans does Connie have and how many credits do they include?” to compare Starter and Scale with their credit options and live monthly/annual prices, plus Enterprise with custom pricing. Top-up pricing is returned separately. See Plans. For clients that read structured tool results, billing carries the call receipt. When billing.replayed is true, it repeats the original receipt without another charge. If a tool also returns its own billing information, its complete output is preserved under result. For example, recovering a saved video can have a 0-credit call receipt while result.billing.credits reports the original generation’s charge. Cached replies keep the same result path. See Credits and Top-ups.

How your AI app sees Connie’s tools

Connie has more than 1,000 tools. Most MCP apps see five general tools and reach the rest through them. ChatGPT also lists common actions directly, including Workflow editors and their variable and account readers. Its Search tools, Get tool and Use tool helpers reach the same public catalogue: a tool does not need to appear in the initial list to run. Search skills and Get skill also read accessible workspace Skills and built-in guides. Apps that support cards in the chat may also see Show image and Show video: You don’t need to call these yourself. Just ask for what you want and your AI app picks the tools.

Tools from a server you connected to Connie

Tools that came from an MCP server connected to your workspace — X Ads, Linear, Notion and the others on Connie’s list — work here too. They’re reached by calling Use tool with the tool’s name, rather than appearing in the default list of five. So your AI app can draft an X campaign through Connie just as Connie’s own chat can. A connected tool’s name is mcp__, the provider, then the tool, joined by double underscores, with any dash in the provider written as an underscore — for example mcp__x_ads__list_ads_accounts. A connected server’s card in Plugins → Accounts lists its tool names on the Endpoints tab. This also works on ChatGPT, through Get tool and Use tool.
  • A server you added yourself won’t turn up in a search. Tools from your own custom MCP servers can be run through Connie’s MCP server with Use tool, by their exact name, but Search tools does not list them — so your AI app has to be told the name. The server’s card in Plugins → Accounts shows its tool names on the Endpoints tab. Connecting that server to your AI app directly is still the easier route if it supports it.

Manage your personal Memory

These tools are available through standard MCP discovery (search-tools, get-tool, then use-tool), not as additional top-level actions in the default tool list. They are also available on ChatGPT. Ask your connected AI app to list, add, edit or delete your Connie memories. The tools are memory-list, memory-create, memory-update and memory-delete. They manage your own entries in Settings → Memory, without needing a Connie conversation. They cannot read or change another member’s memories or Agent memory. memory-list returns active entries, the six category identifiers, and nextOffset; follow nextOffset until it is null to list every entry. Create takes content (up to 400 characters) and category. Update takes an entry’s id and its new content or category, keeping the same entry. Delete takes id and permanently removes the entry. When connected by signing in, you can pass workspaceId to target another workspace you belong to. Ownership stays with your signed-in account. API keys remain scoped to their workspace and must be linked to an execution member. Guests can list memories but cannot make changes through MCP. Disabled tools cannot run; any approval happens in your AI app. Changes appear in Settings and are used in your next conversation that only you read, such as your DM with one of your agents, with Memory on, subject to its 40-entry limit. A room with other people does not use your personal Memory (see Memory and privacy). You can manage saved entries while Memory is off. See Memory.

Good to know

When enabled for your workspace, publish-linkedin-post-with-options supports exact LinkedIn reshares, quoted posts and explicit new-post visibility/commenting controls through authorized MCP apps, including ChatGPT. Use get-post for the source and connector-list-accounts for the acting individual Classic account; clarify ambiguous selections. Ordinary publishing remains write-linkedin-post. These extensions cannot run in Workflows, Records columns or programs. Reuse the original idempotency key on retries, and check LinkedIn before any new request after an uncertain publication. See LinkedIn publishing controls.
  • Who can connect: anyone with a Connie account who belongs to at least one workspace. You don’t need a particular Connie plan. Your AI app may need a paid plan; see each app’s page.
  • Your role applies. Owners, admins and members can read and make changes. Guests can only read through MCP. See Members and roles.
  • Each person connects their own app. Connecting Claude gives your Claude access to your workspaces. Teammates connect their own.
  • Approvals happen in your AI app. Connie doesn’t ask for approval on MCP actions. Your AI app decides whether to ask you first; Claude, ChatGPT, Cursor and VS Code all let you choose per tool. The Connie tool setting Ask before running doesn’t apply to MCP. A tool turned off in Connie can’t run through MCP. See Tool permissions.
  • Tool details describe the call’s price and Chat default. A provider read is reported as operation: read; a zero-credit provider read has risk external_read, while a paid read has risk paid. External billing and account requirements are reported separately. Unknown reach remains external_action. requiresApprovalInChat uses Connie Chat’s policy, including saved workspace settings and connector exceptions; it does not impose Connie Chat approvals on your MCP app. For LinkedIn Ads API Request (Advanced), details without quoteFor report a variable price capped at 1 credit. A concrete GET quote costs 0 credits; writes cost 1 credit.
  • Create and edit Pages without opening Connie. Ask your AI app to create a Page with Markdown, designed blocks or an interactive HTML block. Connie creates a private owning chat and returns links to the Page and chat. Your AI app can list Pages you can access, read them, append sections, replace content, rename them and change their icons. Small inline scripts in HTML blocks work in an isolated frame; they cannot read the host Page, cookies or make network requests, but they can keep each viewer’s inputs and download files such as a CSV. Saving a Page does not publish a public link. A narrow Page search prioritizes matching headings and phrases and returns bounded passages without overlapping text. It includes the Page version and source offsets; an omitted match count describes matching locations outside the returned passages. Search results do not establish a complete review or prove absence. See Page reading and artifacts.
  • Build static apps through MCP. Ask for a static HTML/CSS/JavaScript app. Your AI app can create it, save changed files against a source version and deploy it. Sharing requires a separate request; creating or deploying an app keeps it private. App results include a direct link with the app’s existing permissions. List apps by title, request up to 100 at a time (25 by default), and continue with the returned next offset. Reuse the same operation key when retrying a creation. Scheduled external MCP runs can use these actions without an open chat.
  • Some tools only work inside Connie, for example deleting a Page from its chat, chat follow-ups (“tell me when this finishes”), cancelling a card waiting in a chat, placing calls with the phone dialer and running an AI agent directly. If your AI app says a tool “can’t run via MCP”, it gives the reason. To run an agent from your AI app, put it in a workflow and run the workflow.
  • Read dialing history. Ask about dialing attempts, unanswered calls, voicemail and answer rates by caller or caller number. This read-only tool includes attempts without a processed conversation and never places calls. See Ask about dialing attempts.
  • Check calling sequence progress. Find a sequence by name, see how many distinct people have been attempted and page through people never attempted. Choose a sequence when names are ambiguous. Reads never start or change a sequence. See Sequence progress.
  • Values are saved only as their field holds them. Record changes from your AI app follow the same rules as Connie’s chat: a select value must be one of the field’s options unless your AI app asks to add it as a new one, a date must be a real calendar date, an Email field takes only email addresses, a URL field only web addresses, and a Number or Currency field only numbers. On a record it creates, a value that doesn’t fit is left empty and named in the result. A change to an existing record with such a value changes nothing and says why; text the field already holds doesn’t count against it. See Lists and views.
  • Search results can be saved to a List. When your AI app runs a search through Connie, such as Find People at Company, Connie keeps that result for 30 days, for you, in that workspace. Your AI app can then save the people or companies it found to a List with the search’s own values and where they came from, so they aren’t marked as unverified. Where they came from is recorded with the List, not added as columns; a new List shows the name, Email and Phone and the columns the save filled. It can save several searches together, for example one per company. Only the person (or API key) who ran a search can save its result.
  • Long results can be read to the end. One result is capped at 100,000 bytes. When a web page, a search or a social feed your AI app read through Connie is larger, it gets the readable part first (a page’s text, the first rows of a list) and Connie keeps the whole result for 30 days, for you, in that workspace. Your AI app reads the rest with expand-tool-result, using the result’s _resultRef and the position the result names. Reading on is free: the page is not scraped again and the search is not run or charged again. Results from your own workspace, such as Records, are not kept this way; ask for fewer fields or rows instead.
  • Missing a connected account? If a tool needs an account you haven’t connected yet, such as Gmail or LinkedIn, your AI app gives you a https://connie.ai/connect/... link. It opens Connie on the right connection. The link works for 15 minutes.
  • Staying signed in: your AI app renews its access in the background. If you don’t use it for 30 days, you’ll need to sign in again.
  • Connect my agent stays available in the agent menu, Settings and at https://connie.ai/<workspace-id>/connect-agent, including after an AI app or Slack is connected. It opens once a new workspace’s first run is finished.
  • Opening https://mcp.connie.ai in a browser shows a short “Connie MCP server” page saying it is not a website. That’s expected. Paste the address into your AI app instead.

Record identity and ordering

Record reads return stable Record IDs. Use those IDs for a later read or update; names can repeat, and row positions can change. Search Records follows the captured view for a plain listing from that view. List Records reads the whole Object in its default grid order, with stable ordering across pages. This applies to system and custom Objects and Workflow Records. Record ID filters require a complete UUID with equals or does not equal. Partial matching, such as starts with, is not supported. An incomplete ID or unsupported Record ID comparison returns: “Record ID must be a complete UUID. Partial matching isn’t supported. Use “equals” with the full Record ID.” The legacy Get a Cell, Clear a Cell, and Clear a Row tools take a position in the unfiltered Object’s default order. Do not pass them a position from a List or a sorted or filtered view. For a Record found there, use its exact ID with Get Record or Update Record instead. A Standard workspace whose trial (started before November 2026) has ended needs a subscription for paid tool calls, even if credits remain. The refusal includes reason: trial_ended; remaining credits are kept. Owners and admins receive the subscription recovery step, and other members are told to ask one. ChatGPT receives the public pricing page. Beta / credit-only and Internal workspaces retain their access exceptions. See Trial end. A workspace whose free credits are used up is refused with reason: free_credits_spent until an owner or admin chooses a plan.

Received LinkedIn connection invitations

When released in your workspace’s catalog, discover List LinkedIn Invitations (inbox-list-linkedin-invitations) and Accept LinkedIn Invitation / Decline LinkedIn Invitation (inbox-accept-linkedin-invitation, inbox-decline-linkedin-invitation). Use connector-list-accounts to choose whose connected account to use; account access and non-guest workspace membership apply to the reader and decisions. Keep the selected account’s opaque accountRef in the host envelope when required, separate from the exact accountId in the tool input. The reader returns at most 20 received requests per call, with readable sender information and exact accountId, requestId and offset for a decision. Continue with nextOffset when present, including short or filtered-empty pages. Clarify ambiguous senders or accounts. Accept changes the relationship; Decline refuses the incoming invitation with destructive confirmation and never withdraws an outgoing invitation. For Decline, set confirmed: true only after confirming the exact sender request and selected account. Both recheck current received-request eligibility before mutation. Reuse the original idempotency key on transport retries. Never automatically replay an uncertain decision or one accepted with local refresh pending, including through the opposite action. Refresh stale requests without inferring whether they were accepted or declined. Failed lookups stay errors. Shared policy and billing apply. Agent chat and authorized general MCP/ChatGPT only; no Workflow Actions/Agents, Records columns or program execution, including listing. See received invitations.

LinkedIn Recruiter projects and pipeline candidates

When released in your workspace’s catalog, discover list-recruiter-projects, get-recruiter-project, create-recruiter-project, edit-recruiter-project, list-recruiter-pipeline-candidates and save-recruiter-pipeline-candidate through search/get-tool and the supported invocation path. They use an exact accessible connected Recruiter account and its selected contract. Keep the opaque accountRef in the host envelope, separate from the exact accountId in tool input. Copy account/project bindings and candidate/stage IDs from prerequisite reads; clarify ambiguity. Pipeline listing is a read even though its provider request uses POST. These tools exclude Workflows, Records columns and programs. Shared access, billing and approvals remain; never automatically repeat an uncertain write. See Recruiter projects and pipelines for fields, pagination and duplicate handling. When released in your workspace catalog, discover search-recruiter-talent-pool using search/get-tool, then invoke it directly by ID or through the supported tool wrapper. It is absent from ChatGPT’s compact initial list and available in the general full list when enabled. First use connector-list-accounts and list-recruiter-projects / get-recruiter-project; keep the opaque accountRef in the host envelope and copy the exact account/project binding into input. This member-only reader resolves the project’s active Recruiter search channel internally and supports a strict filter subset. Location labels resolve internally without guessed IDs. Copy the bound nextCursor with identical filters/limit. Candidate User Profile and candidate IDs remain distinct; a non-null actionIdentifier can feed a separately authorized existing save action. Searching does not save, contact, create or write Records. Shared access, pricing and cancellation apply. No Workflow, Records-column or program access. See Talent-pool search. schedule-linkedin-post saves a post to be published later from one connected account or managed Page; get-scheduled-linkedin-posts, update-scheduled-linkedin-post and cancel-scheduled-linkedin-post read, change and cancel it until it goes out, and compare-linkedin-post-draft measures a draft against the accounts’ recent and scheduled posts. Each post is read back from LinkedIn after publishing and recorded live with its link or not published with the reason. See Schedule LinkedIn posts. When enabled, write-linkedin-post can also publish as a managed company Page. Select the connected account, read list-linkedin-managed-pages, and pass its exact accountId/id/name binding as company; keep opaque accountRef in the MCP envelope. Omission preserves personal publishing; an invalid or revoked Page never falls back to personal posting. The result distinguishes the actual posting identity from the authorizing individual account. See company publishing for limits and safe recovery after uncertain publication. search-lists returns viewUrl for accessible Records Lists. Use this workspace-relative URL as the destination of a Markdown link in a Page. Folders and artifact tables omit this field. Page body writes also turn bare current-workspace List paths into links, while preserving existing links and code examples. Opening a source still requires access to that source.