Skip to main content
Find your problem below. Error messages are quoted exactly as your AI app shows them, so you can search this page for the text you see. Where to check your connections: left sidebar → Plugins → Accounts → My accounts · https://connie.ai/<workspace-id>/tools?tab=my-accounts

Signing in and connecting

Why: an unfinished first run keeps Settings closed, including Developers → API keys. A direct https://connie.ai/<workspace-id>/settings?tab=developers link follows the same check. Fix: answer the first run’s two questions (your website or a description, and your phone number), then reopen API keys. New workspaces use free credits, with no card or trial. Plugins, opened Pages and Billing (?tab=plans) remain available during setup. If setup is already complete, reload the same workspace and contact support if the redirect continues. See Settings access during setup.The current API-key guide is API keys. Older /mcp/api-key links redirect there; they do not provide another way to unlock Settings.
Full message: “Your Connie sign-in expired. Your MCP client normally renews it automatically; if this keeps happening, disconnect Connie in your client and connect it again.”Why: the app wasn’t used for 30 days, or it couldn’t renew its sign-in. Fix: authenticate again in the app. In Claude or ChatGPT, open Connie and click Connect. In Claude Code, run /mcp → connie → Authenticate. In Codex, run codex mcp login connie --scopes mcp:full. In Cursor, click Authenticate.
Full message: “Your Connie sign-in was revoked, for example after access was removed in Connie. Disconnect Connie in your MCP client and connect it again to sign in.”Why: the app’s access was revoked in Plugins → Accounts → My accounts. Fix: remove Connie in your AI app, add it again and click Allow access.
Why: the app was set up with an old or wrong server address. Fix: disconnect Connie in your app and add it again with exactly https://mcp.connie.ai.
Why: the same stored sign-in was used from two places, for example a restored backup or two installs sharing copies of one login. Connie ends a reused sign-in to keep your account safe. Fix: sign in again from the app. Don’t copy an app’s stored login to another install.
Why: the Connie access screen was opened more than 10 minutes after the app started the sign-in, you’re signed in to Connie as a different person than the one who started it, or your account has no active workspace. Fix: start the connection again from the app, and finish it straight away with the right Connie account.
Why: the app’s registration with Connie is out of date, or the app didn’t register itself with Connie (for example, a Claude custom connector set to use Claude’s published identity or your own OAuth client). Fix: remove Connie from the app and add it again, so it registers afresh. In a Claude custom connector, choose Register automatically under OAuth client.
Why: you’re browsing Connie as a guest visitor, not signed in to your account. Fix: sign in with your Connie account and the connection continues.
Sign up or log in to Connie, then continue the connection flow. New workspaces start with 1,000 free credits, with no card or trial. Browsing anonymously does not authorize an AI app to use your workspace. See Sign up and sign in.
Full message: “These Connie credentials belong to someone trying Connie without an account, and only signed-up accounts can use API keys. Nothing was run. Sign up in Connie, then copy your personal key from https://connie.ai/<workspace-id>/settings?tab=developers and update this MCP server’s settings.” The response code is anonymous_account.Why: the app is set up with an API key made while you were trying Connie without an account, and only signed-up accounts can use API keys. Fix: sign up, finish your workspace’s first run, then open Settings → API keys, copy your Personal key and put it in the app’s settings. Or remove the key and connect with sign-in (OAuth) instead. See API keys.
Full message: “This older workspace API key isn’t linked to a current member of the workspace, so Connie refused it. Nothing was run. Copy your personal key from https://connie.ai/<workspace-id>/settings?tab=developers and update this MCP server’s settings.” The response code is legacy_api_key_no_member.Why: the app uses the workspace’s original single API key, and the person that key acts for (or, if it names no one, the person who created the workspace) is no longer a member of the workspace. Fix: open Settings → API keys, copy your Personal key and put it in the app’s settings. Or remove the key and connect with sign-in (OAuth) instead. See API keys.
Why: the app may retry sign-in automatically. Connie’s access screen shows ‘You recently revoked [App]’ with instructions for that app. Fix: click Cancel, remove or disable Connie in the app’s settings, fully quit and reopen the app, then revoke in Connie. Codex users can stop a running terminal sign-in with Ctrl+C and run codex mcp remove connie. Cursor users can remove Connie in its MCP settings or from ~/.cursor/mcp.json. If you intentionally started sign-in to reconnect, you can allow access on this screen.
Why: affected Codex CLI versions retry automatic scope discovery after cancellation. Connie returns access_denied and grants no access. Fix: press Ctrl+C in the terminal to stop the sign-in. When you want to connect, run codex mcp login connie --scopes mcp:full, also available as Sign-in command in Connie’s Codex Manual guide. Explicitly naming the scope avoids that retry.
Fix: leave them empty, or choose automatic registration. Connie registers your app by itself and doesn’t use client secrets.
That’s expected. https://mcp.connie.ai is an address for AI apps, not a web page. Paste it into your AI app’s connector or MCP settings.
Why: the app is set up to send an API key header but none was found, or the header is wrong. Fix: we recommend removing the header and signing in with OAuth instead. If you do use a key, see API keys for the right format.

Workspaces and permissions

Full message ends with: “Ask one of its admins to invite you, or pick one of your own workspaces with list-my-workspaces.”Why: the app tried a workspace you don’t belong to or were removed from, or a workspace ID in the server URL is wrong. Fix: ask your AI app to list your Connie workspaces and use one of them. To join another workspace, ask its admin for an invite.
Why: the app passed a workspace name where Connie needs its ID. Fix: ask your AI app to “list my Connie workspaces” first, then repeat your request.
Why: that’s your Default workspace, and reconnecting keeps it. Fix: name the workspace in your request, or change the default: revoke the app in Plugins → Accounts → My accounts, then connect again. See Manage access.
Why: you’re a guest in that workspace. Fix: ask a workspace admin for member access.
Why: an API key belongs to one workspace. Fix: use that workspace’s own key, or connect with sign-in (OAuth), which reaches all your workspaces.
Why: the workspace was deleted. Fix: to restore it, email support@connie.ai within 7 days of the deletion (there’s no restore button in Connie). Otherwise, use another workspace.

Generated images

If a generated image appears only as a link, ask the app to display the existing result with Show image. Start a new chat or reconnect Connie if the app has an older tool list. Apps without card support still receive a link to the image. To reopen a previous result, ask for get-generation with its saved generationId, then Show image. You can also open the generating tool in Plugins (its History tab) and use Download. Neither action starts another paid generation. If the provider file has expired, its download may no longer be available.

Tools, credits and limits

Full message: “Slack did not return a message for this channel and timestamp. Check that the channel and timestamp came from the same message and that the connected account can access it. The message may have been deleted or removed by the workspace’s retention policy.”Why: Slack’s history lookup completed but did not return the requested message. This does not prove that the channel or timestamp was entered incorrectly. Fix: check the message in Slack, the selected account’s access, and that the channel and timestamp belong to that message. Read the available channel history to choose a current message if the original is gone. A message with no replies is returned as a one-message thread when Slack can find it.If the message instead says “Slack could not verify this thread because its history lookup failed. Try again later; if it keeps failing, check the connected account’s access.”, Connie could not establish whether the message exists. Retry later or check the connection’s access; do not assume the message was deleted or the timestamp was wrong.
Full message: “Workflow <id> not found in this workspace (wrong id, or it may have been deleted). Nothing was run. Use workflow-list to find the right Workflow id.”Why: the ID passed to run a Workflow doesn’t match a Workflow in the current workspace. The response carries RESOURCE_NOT_FOUND and noSideEffects: true; no Workflow Actions run.Fix: ask your AI app to list Workflows in the intended workspace with workflow-list, then use the correct ID. Repeating the same missing ID won’t start a run. See Test and run a Workflow.
Why: the initial tool list is compact. Most AI apps see five core tools; ChatGPT also lists common actions directly. All can reach other public tools through Search tools, Get tool and Use tool, so absence from that list does not mean a tool is blocked. Workflow editing, message sending, media tools and connected-server tools are available on ChatGPT with the same workspace permissions. A few tools only work inside Connie (for example deleting a Page from its chat, chat follow-ups, cancelling a card waiting in a chat, placing calls with the phone dialer and running an agent directly); the app tells you why. Fix: just ask for what you want. Creating, listing, reading and editing Pages, and managing static apps, work through authenticated MCP without an open Connie chat. If an existing Page is unavailable, check that you can open its owning chat in the same workspace. To run an agent, put it in a workflow and ask the app to run the workflow. Reading dial attempt history is supported; see Calls and the dialer. See Use Connie in your AI app.
Why: the tool is turned off in Connie. Fix: an admin can turn it back on. See Tool permissions.
Why: approvals for MCP happen in your AI app, not in Connie. The Connie setting Ask before running doesn’t apply to MCP. Fix: set Connie’s tools to need approval in your AI app, for example Needs approval in Claude.
Why: the action needs an account such as Gmail or LinkedIn that isn’t connected in this workspace. Fix: open the https://connie.ai/connect/... link the app gives you within 15 minutes and connect the account. Connie puts that link in the message itself, so every AI app can show it, including apps that only pass the message text to their model. If you see ‘Connection link unavailable’, or the app gave you no link, ask your AI app for a new one. When several accounts are connected and none is the default, the message lists them so the app can ask which one to use. See Connect accounts.
The message reads “This id is retired. Use ”…” instead: it is the current tool for this.” or “This tool is retired: its catalogue entry is switched off. Find what replaced it with search-tools.”Why: your AI app called a tool by an old name, usually from an earlier chat or a saved instruction. Nothing ran and nothing was charged. Fix: ask the app to try again; it now has the current tool’s name, or can find it with Search tools. If a saved prompt or skill of yours names the old tool, update it.
The message starts “Not enough Connie credits…” (or “Your guest credit budget… is used up”) and ends “Nothing was charged.”Why: the workspace doesn’t have enough credits, or your guest budget is used up. Fix: owners and admins follow the link the app gives to add credits or change plan. Members and guests ask an owner or admin. See Credits and Top-ups.
Why: a Find Leads or Find Companies call has no search criteria. The tool returns INVALID_INPUT before searching, importing or charging for the search. A List name, result limit or preview mode does not count as a search filter.Find Leads: “At least one search filter is required — pass person/company profile filters, funding filters, or investors.”Find Companies: “At least one search filter is required — pass company profile, revenue, founded-year, funding, or investor filters.”Fix: ask your AI app to include at least one filter that matches your request, such as a job title for people or an industry for companies, then try again. See Lead Finder.
Why: Lead Finder imports through an external AI app cannot include email or phone lookups because those have separate variable credit costs. The rejected call imports no leads and runs no contact enrichment.Fix: ask your AI app to import the confirmed leads with findEmail and findPhone omitted or set to false, keeping the same confirmed filters, destination, quantity and preview token. Then request email or phone enrichment separately through the dedicated contact-detail tools, with their own pricing and confirmation. See Enrichment.
A prompt edit can return INVALID_INPUT with “The text to replace was not found in “task”, so nothing was changed. Read the current value and quote it exactly, including whitespace and casing.”Why: object-update-field uses configReplace to match saved text exactly. The saved prompt may use a tool’s display name instead of the internal ID originally submitted. If any replacement cannot be applied, none of that call’s edits are saved.Fix: ask your AI app to read the current Field or Workflow Action configuration, then retry with the exact saved text. For a Workflow Action’s prompt, it can use workflow-get with the Workflow ID, Action alias and fieldKey: "task". If a match appears more than once, quote more surrounding text, or use all: true only when you want every occurrence replaced.
Updating a task can return “Notes / Description can hold up to 5,000 characters including formatting, and this text needs 5,038. Shorten it by at least 38 characters and try again. Nothing was updated.” (with the numbers for your text). Creating a task returns the same message ending in “Nothing was created.”, and saving a Record’s notes returns it starting with “Notes” and ending in “Nothing was saved.”Why: a Record’s notes can hold up to 5,000 characters. For a task’s Notes / Description, paragraphs and line breaks count toward that limit, so text slightly under 5,000 characters can still be too long; for saved notes, any HTML tags count. Nothing is changed: the task isn’t created or updated (including any other Fields in the same request), and existing notes stay as they were.Fix: ask your AI app to shorten the notes by at least the number of characters shown, then try again.
Why: an array plan requires a snapshot of the whole source List, including its Record count and membership hash. The tool returns INVALID_INPUT with “arrayPlan currently requires selection:all_existing so the source count and membership hash are complete.” when that selection is missing. No plan is created by this refused call.Fix: ask your AI app to use selection: "all_existing" only if you intend to plan for the whole source List. This array-plan path cannot plan for a subset; keep your intended scope rather than widening it to clear the error.
Why: object-native-actions needs desiredOutputs, an exact actionId or actionNames, or an arrayPlan. A call without these returns INVALID_INPUT with “Supply desiredOutputs, one exact actionId/actionName, or an arrayPlan for an entire-List assessment.” The unsupported enrichments parameter does not select outputs. No assessment or plan is created by this refused call.Fix: ask your AI app to use the supported parameter that matches your request, keeping the same intended Records. actionNames accepts the registered name, its advertised canonical name, or a retained former name for an available Action; for example, “Find Email (from LinkedIn URL)” and “Find Email (LinkedIn URL)” refer to the same Action. If a name is unavailable or ambiguous, use operation: "search" and copy the exact returned Action ID. Connie does not substitute another provider to clear the error.
Why: more than 300 requests a minute, or 3,000 an hour, from one sign-in in a workspace without a paid plan (3,000 a minute and 60,000 an hour on a paid plan). Nothing was run. Fix: wait the number of seconds in the message and try again, or choose a plan for much higher limits. See Errors and limits.
Why: the connection was idle for 30 minutes. Fix: most apps reconnect by themselves. If yours doesn’t, reconnect Connie or restart the app.
Why: your AI app couldn’t connect to Connie. Before sending an MCP POST request, Connie waits up to 10 seconds for its tool server to become ready, checking every second. If it becomes ready, the held request proceeds automatically. The wait stops if your app disconnects. Brief connection failures when reopening the notification stream are retried automatically once. Small requests that only read, such as listing tools, are also retried once when the connection drops. Other small requests receive one automatic retry only when the connection times out before Connie receives the request. Actions that may already have started are not automatically repeated by this connection retry. Fix: if this message still appears, wait a minute and try again.
Why: Connie could not read the agent’s work status. Connie can retry a temporary read failure once while time remains in the wait. This status check does not stop or restart the agent’s work, and an unavailable status does not mean that the work failed.Fix: ask your AI app to check again with agent-task-wait and the same taskId. Do not send the original work request again. A normal wait that ends while the agent is working returns pending with waitTimedOut; a status that could not be retrieved returns an error instead. Setting timeoutSeconds to 0 makes one status read without an automatic retry.
Why: the LinkedIn data provider could not complete a lookup, even after Connie’s automatic attempts. The failed lookup is returned to your AI app as a tool error.Fix: “Please try again in a minute. You do not need to change your filters or add credits. If it keeps happening, contact support.” Your AI app can retry this read-only lookup once after a minute with the same tool and inputs. If that attempt also fails, the app should show you the message and stop retrying. This retry guidance applies to data lookups with reference [5104]; it does not authorize repeating a send or another action whose outcome is uncertain.
Why: Get Profile and related LinkedIn data lookups share data source capacity across Connie. A streaming MCP call waits automatically for up to 15 minutes of total capacity waiting, including waits between pages. If your AI app supports progress, it shows Waiting for data source capacity. A plain JSON call may return this message sooner. Cancellation, disconnection or server shutdown can also end a wait earlier.Fix: follow the retry delay returned by Connie, then retry with the same tool and inputs and the same idempotencyKey if you supplied one. Saved lookup progress remains available for supported resumptions. You do not need to change your filters or add credits to free data source capacity. A request already sent to the data source may still finish after you disconnect.
Why: some web data tools collect more than one call can wait for, so the tool answers straight away with a resultId and the collection carries on in the background. Large collections can take 20 minutes or more. Fix: let your AI app fetch the result with that resultId — it can keep checking until the collection answers, and the result can be fetched for 24 hours. Don’t ask for the original tool again: that starts a second collection and charges for it. If the collection fails, the credits for the call that started it are returned.
Fix: try again. If it keeps happening, email support@connie.ai with the reference, the AI app you use and the time it happened.

App-specific problems

On Team and Enterprise, only owners add connectors; members connect them or click Request. On Free, you can have one custom connector. Use the directory listing instead. See Claude.
That file is for local servers only. Remove the entry and add Connie from the directory or Customize → Connectors. See Claude.
Why: your chat still uses the Connie tool list from before 11 October 2026, when every action opened a video player. Fix: start a new chat. If it still shows, disconnect and reconnect Connie in Customize → Connectors. Now the player opens only when you ask to watch a video Connie returned.
Ask Connie to run the tool or skill that returned the videos again, then show the videos from that fresh result. You can also open the original video links. If Connie cannot verify one video’s access key, the whole request may be refused; that does not necessarily mean the key expired. If a fresh result still fails, contact support.
Open Connie’s plugin page and click Install plugin. Personal and Business accounts follow the same steps. If installation is unavailable, report the message shown to support. See ChatGPT.
In VS Code’s own mcp.json, the top-level key must be servers with "type": "http". Start the server from MCP: List Servers. See VS Code.
Use httpUrl, not url, in settings.json. See Other apps.
Why: the sidebar no longer has that card. Fix: choose Connect my agent in the agent menu (click the agent card at the top of the sidebar) or Settings → Developers. While a new workspace’s first run is unfinished, Settings and https://connie.ai/<workspace-id>/connect-agent send you back to it; reopen them after finishing it.
Why: your browser blocked the new tab. Fix: allow pop-ups for connie.ai, or open claude.ai/directory/connie yourself and click Connect to Claude.

Helping support investigate a tool call

When reporting a problem, include the tool name, the time it happened and the request ID if your AI app exposes it. Connie keeps staff-only MCP request logs for 90 days. For authenticated tool calls, these include the parameters your app sent, the outcome and any error. Recognizable credentials and inline file data are removed; large or deeply nested parameters may be shortened. Older requests and requests rejected before authentication may have no saved parameters. Tool response bodies are not included in these request logs.

A workflow approval is closed after Stop

Stopping a workflow run closes its pending approvals. Pausing, archiving or deleting the workflow also closes affected requests. An old request cannot authorize a rerun. Run the workflow again to request fresh approval when it reaches a step with Require approval enabled. A closed request returns “This approval is no longer open, so your decision was not recorded.” If cancellation could not be sent, approval requests remain closed while cancellation delivery is retried. A cancellation error can say “Approval requests are closed. Retry cancellation if the run is still active.”