Connectors let your agents attach files from Google Drive, receive work from Linear and coding agents, and act inside 1,500+ apps through connected tools, keys, and your own MCP servers.
Connectors come in three classes, each with its own page under Connectors:
Important: connecting a service does not by itself give agents access in a chat. You turn tools on per chat (or per AI employee) with the Tools chip in the editor, and actions above your approval threshold wait for your approval.
Open Connectors from the sidebar, or go to cellcog.ai/connectors. The hub shows the three classes, a search box for the tools catalog, and Your connections, a list of everything you have connected with a Manage link on each.
Connect Google Drive to attach files without downloading them first.
Connecting:
Using Google Drive:
Disconnecting: click Disconnect on the Google Drive page.
Delegate lets you bring CellCog into the agents and platforms your team already uses. Instead of coming to CellCog, CellCog comes to you: assign a task from within Linear, Cursor, Claude Code, or OpenClaw, and results flow back where you were.
| Platform | How it works |
|---|---|
| OpenClaw | Generate an API key, hand it to your OpenClaw agent, and it delegates through the CellCog Python SDK and ClawHub skills |
| Linear | Assign an issue to @CellCog or @-mention it in a comment; the agent session streams back into the issue |
| Cursor | Install the CellCog plugin; /cellcog and the skill commands appear in the command menu |
| Claude Code | Install the CellCog plugin; slash commands light up CellCog inline |
| Other Agents | Custom scripts, direct HTTP callers, or any agent framework that can make API requests, with the same API key |
Open MCP / API in the left sidebar and pick your agent to see every way it connects. Each platform also has its own page there, with setup steps, connection status, and the number of tasks delegated so far:
Tools let agents perform actions in external services during chats. The catalog covers 1,500+ apps with 20,000+ tools: Gmail, Slack, Linear, Jira, Notion, HubSpot, Salesforce, Shopify, Stripe, Figma, Canva, Google Analytics, and many more.
Each tool's page shows what agents can do with it before you authorize. Some tools are built in and need no connection at all; they are ready in any chat where Tools is enabled.
The Tools chip in the editor has three states:
The chip shows what is active (all of your connections, or how many you picked). You can also search the catalog and connect a new service from inside the chip. The selection is saved with the chat.
AI employees hold their own Tools selection, set when you hire them and editable afterwards, so each employee works with only the providers its role needs.
Every tool call is classified before it runs: safe (reading non-private data), moderate (reading private data such as full email content or payment history), or dangerous (anything that changes data or acts on your behalf, such as sending an email or posting publicly). On the Tools page, Auto-approve tool calls up to sets your account-wide level: None, Safe, Moderate, or Dangerous. Calls above the level pause and wait for your approval in the chat. Tools from your own MCP servers are always treated as at least moderate.
For the full approval model, including per-chat overrides, see the Approvals and Safety Guide.
If a service revokes or expires its authorization, the tool's page shows that the connection has expired and offers Reconnect. Agents get a clear message instead of failing silently. When a service adds capabilities that need new permissions, the connection keeps working and the page recommends a reconnect to enable them.
Disconnecting removes the connection and its credentials from your account. Agents lose that service immediately.
Many connected tools can also raise events, such as a new email in Gmail, a new calendar invite, a new GitHub issue or a failed Stripe payment. A trigger turns one of those events into a wake for an AI employee. Your employees find triggers alongside tools and may set up the ones their job clearly needs; for anything less clear they send you a trigger link (cellcog.ai/connectors/triggers/...) that opens as a card with an Enable button. If the tool is not connected yet, the card asks you to connect it first. Opening a link never enables anything by itself. Each employee's triggers are listed and managed in the Wake app on its Employee Computer. Triggers need a standing AI employee to wake; regular chats do not use them. When no app event fits, an employee can build a watcher: a scheduled check that reads your connected tools the way the employee would, and wakes it only when something worth acting on shows up. Watchers only read; see the AI Employees Guide for details.
Some services are not in the tools catalog, or an agent's scripts need a raw credential directly. The Keys page at Connectors, then Tools, then Keys (cellcog.ai/connectors/keys) is the home for these.
SERVICE_KEYTYPE; the page previews the exact name as you typeEach row shows the service, the variable names your agents see, where the key came from ("Added when you connected X" or "Added by you"), and two capability marks: Actions (agents can call the service's tools directly) and Scripts (agents can use the key in code they run for you). If a key you added matches a service in the catalog, Enable actions turns on direct tool calls using that same key.
Tools are connections you authorize with a click; agents call their actions directly. Keys are credentials you paste; agents' own scripts read them from the environment. If a service is not in the catalog, add its credential on the Keys page.
Beyond the tools catalog and Keys, you can plug in any remote MCP server: your company's internal tools, a self-hosted service, or any vendor that exposes an MCP endpoint. MCP (Model Context Protocol) is the open standard AI platforms use to connect to tool servers.
Go to Connectors, then Tools, then MCP servers (cellcog.ai/connectors/mcps). Enter a Name, the Server URL (a public https MCP endpoint), and pick the Authentication:
Click Add server. CellCog connects, discovers the server's tools, and lists them on the server's card. Tools are kept in sync daily; use the refresh button on the card to re-discover them anytime.
The server's tools work like catalog tools: your agents find them in tool search and call them in any chat where Tools is enabled, behind your approval threshold. Calls to your own MCP servers are always treated as at least moderate risk, since a third-party server's tools are not vetted by CellCog. Your MCP servers are private to your account.
Each card shows a status: Connected, Connecting…, Sync failed (the URL is unreachable, not a valid MCP endpoint, or exposes too many tools), or Key rejected (the API key was refused). OAuth servers show Sign in required, Finishing sign-in…, or Sign-in failed; whenever an OAuth server is not connected, its card carries a Sign in button, so reconnecting after an expired sign-in is one click. Any card also has Disable (keep the server but hide its tools), Rename, Change URL or auth (replace), and Remove.
Tools are the built-in catalog with one-click connect. Keys are raw credentials your agents' scripts read from the environment. MCP servers are your own tool servers, discovered and called like catalog tools. If a service is not in the catalog: use Keys when you have an API key and the agent can script against the service's API; use MCP servers when the service (or your team) exposes an MCP endpoint, so agents get schema-validated tools instead of raw scripting.
The Logins page at Connectors, then Logins (cellcog.ai/connectors/logins) keeps website usernames and passwords for your AI employees' cloud browsers, so an employee can log in to a site by itself.
example.com), the username, and the password, or generate a new oneAn employee that needs a login can send you a link that opens this form with the site already filled in.
Click Reveal and confirm it is you by signing in again. You see the password, and the current 2FA code if the login has a 2FA key. You can reveal up to 10 times an hour.
Keys are credentials your agents' scripts read from their workspace environment. Logins are website passwords that are only ever entered into a web page by the employee's cloud browser, and never shown to agents.
No. Connect once, then decide per chat with the Tools chip: All, Choose, or Off.
No. Tools must be enabled in the chat, and calls above your approval level wait for you.
The agent loses access immediately. Ongoing operations may fail.
OAuth tokens and the credentials for catalog tools and MCP servers are held encrypted by our tools provider. Keys you add on the Keys page are stored encrypted and never shown again. Saved logins are stored encrypted, and only you can reveal one, after signing in again.
No. An employee's cloud browser enters a saved password straight into the site's page. The value never appears in the employee's scripts, results, or messages.
Yes. Tool calls, their summaries, and their results appear in the chat.
The tool's page shows the connection as expired and offers Reconnect. Agents are told the connection needs reconnecting instead of failing silently.
They draw from the same connections and keys, but each employee has its own Tools selection, so you can limit an employee to the providers its role needs.
Connections and keys belong to you, not to an organization, so the tools you connect are available to your chats and AI employees in every organization you belong to.