Connect CellCog to Claude, Cursor or any MCP client. Sign in once with your CellCog account and brief finished work from inside the session you already use, no API key needed.
CellCog runs a remote MCP server at:
texthttps://cellcog.ai/mcp
MCP (Model Context Protocol) is the open standard AI assistants use to reach outside tools. Once you connect CellCog, the assistant you already work in (Claude web, Claude Desktop including its Cowork sessions, Claude Code, Cursor, or any client that speaks MCP) can hand CellCog a brief and get finished work back as links, without leaving the session.
What CellCog does with the brief: an agent takes it as a job and returns the finished result, not a single model call. The script, the asset and a check of the result come back together. That covers videos with dialogue, images, voiceovers and cloned voices, music and sound, podcasts, talking avatars, 3D models, decks, PDFs, spreadsheets, dashboards and web apps, diagrams, research with sources, data analysis and code, plus edits such as upscaling, background removal, outpainting and dubbing in the same request.
Both let another agent use your CellCog account. Pick by where you work:
| MCP connector (this guide) | API key + SDK or plugin | |
|---|---|---|
| Sign-in | Your CellCog account, in the browser, once | A key you generate and store |
| Works in Claude web, Claude Desktop, Cowork | Yes. These surfaces cannot keep a package or an environment variable between sessions; a connector is the one thing that persists | No |
| Works in Claude Code, Cursor, OpenClaw scripts | Yes | Yes |
| Managed from | Settings, then API keys, then Connected apps | Settings, then API keys |
| Counts toward the 10-key limit | No | Yes |
The Coding Agent Plugin and the Python SDK (API Keys Guide) keep working exactly as before. The connector adds the surfaces they could not reach.
New accounts see CellCog MCP as one of the choices on their first screen after sign-up: it shows the address and the two ways to add it, and turns into "CellCog MCP is connected" the moment the connection lands. Connecting from Claude first works the same way; the first screen then knows you have done it.
https://cellcog.ai/mcp and confirmCellCog now appears in your connectors with its thirteen tools. By default Claude asks you before each tool call ("Needs approval"); the dropdown next to the tool group switches them to always allow. That switch is Claude's and is set per user; CellCog does not change it.
The seven tools that only read (status, results, account, recent chats, and listing, showing and reading your AI employees) are marked read-only for the client. The six that act (create a chat, send a message, open an upload window, stop a chat, hire an AI employee, message an AI employee) are marked additive and never destructive. Clients that honor these hints may run the read-only tools without asking each time.
This is Use CellCog from Claude Code: Claude Code hands work to CellCog. For the other direction, your CellCog agents starting Claude Code, see Use Claude Code as a subagent in the Subagents Guide.
In a terminal:
bashclaude mcp add --transport http cellcog https://cellcog.ai/mcp
The connector joins the current session; /mcp lists it. The first use opens the CellCog consent page in your browser. Approve once; the sign-in persists for future sessions.
If you already use the CellCog plugin, version 2.3.0 and later includes the connector. Add the marketplace and install the plugin:
bashclaude plugin marketplace add CellCog/cellcog-plugin
Then install the CellCog plugin from that marketplace as described in the Coding Agent Plugin guide. You get the connector and the skills in one step.
In a terminal, add CellCog and sign in:
bashcodex mcp add cellcog --url https://cellcog.ai/mcp codex mcp login cellcog
The login opens the CellCog consent page in your browser; click Approve. Type /mcp in Codex to see CellCog listed. In the Codex editor extension: open the gear menu, then MCP servers, then Add server, choose Streamable HTTP, paste https://cellcog.ai/mcp, then Authenticate.
Add a remote (streamable HTTP) MCP server and paste https://cellcog.ai/mcp. Every client that supports the current MCP authorization flow discovers the CellCog sign-in on its own: the first use opens the consent page, you approve, and the connection is kept.
Clients that only support local (stdio) servers cannot connect; use the plugin or the SDK there.
On a Claude Team or Enterprise plan an Owner can distribute the CellCog plugin, which carries the connector, to the whole organization:
CellCog/cellcog-plugin as a github source, or upload the plugin ZIP directlyEvery member then sees the CellCog connector and signs in with their own CellCog account on first use. Each person keeps their own chats, memory and billing; nothing is shared through one credential. Claude Code fleets can also carry the server through Anthropic's managed settings; see Anthropic's documentation for that path.
If your organization uses CellCog's shared credit pool, connector chats draw from it like any other chat by an enrolled member. See the Organization Guide.
The consent page lists three permissions:
Behind them are thirteen tools. Claude lists them by the name in the first column; the second column is what a client log or a script shows. The names group themselves: chat tools start with cellcog_chat_, AI employee tools with cellcog_employee_:
| Shown as | Tool name | What it does |
|---|---|---|
| Chat: create | cellcog_chat_create | Starts a CellCog chat from a brief. Choose the agent (agent, creative, team) and the depth (flash, core, max); the agent's default applies when you leave depth out. Attach files you uploaded with Upload files (up to 100 MB each), or public https links up to 25 MB each. Returns the chat id and a link right away; the work runs in the background |
| Chat: send message | cellcog_chat_send | Sends a follow-up into an existing chat, in context: iterate, answer a question the agent asked, or add an attachment |
| Chat: status | cellcog_chat_status | A cheap status check: running, needs_input, done or idle, plus the agent's latest progress note |
| Chat: stop | cellcog_chat_stop | Stops a chat that is running in the wrong direction. The agent finishes its current step, stops and waits, and keeps everything it made so far; redirect it with Chat: send message or collect what exists with Get result. Does nothing on a chat that is not running. An AI employee has no stop: message it, or pause it in the web app |
| Chat: list | cellcog_chat_list | Recent chats with their status, connector-created ones first. Filter by the chat's name |
| AI employee: hire | cellcog_employee_hire | Hires a new AI employee from a brief: the role in plain words, as you would type it on the hire page, and optionally a name. See AI Employees Through the Connector below |
| AI employee: list | cellcog_employee_list | The AI employees you can instruct: the ones you own and the ones a teammate gave you team access to in your organization, each with its id, name, title, email, whether you instruct it as its owner or on the owner's behalf, its status and its current shift |
| AI employee: profile and shifts | cellcog_employee_get | Who one AI employee is and its list of shifts, newest first, with each shift's number, dates and how it ended |
| AI employee: send message | cellcog_employee_message | Sends a message into the AI employee's current shift, exactly as if you typed it in its chat, and says which shift it landed in. Attachments work as on Chat: create |
| AI employee: read a shift | cellcog_employee_shift | Every message of one shift, by its number; without a number, the latest shift |
| Get result (chat or AI employee) | cellcog_get_result | Answers at once. While the chat is still running it returns the agent's progress notes since the last call, each with a time, so the assistant can tell you what CellCog is doing; call it again about once a minute (video jobs take 10 to 40 minutes). When the job is done it returns every reply the agent wrote since your last message, the files, what that reply cost next to the running total for the chat (or the AI employee's shift), and a one-line summary of what was made. In clients that render inline views the result shows video and audio players, images and the first pages of any PDF; a spreadsheet shows a per-sheet table of live formulas against typed-in numbers, whether every formula carries a computed value, and its cell values and formulas as text; markdown in the message, tables included, renders formatted; every file has a Download link valid for 30 days, and every file except a video (which plays inline) also has Open. For an AI employee, give its id instead of a chat: the result follows the employee's current shift, or reads the shift number you give |
| Upload files | cellcog_upload | Opens a 24-hour upload window for files on your computer, up to 10 files of 100 MB each, shown inline in Claude web and Desktop or as a page on cellcog.ai. The agent then attaches them to a chat |
| Account | cellcog_account | The connected account: email, organization, plan, credit balance and parallel-chat slots. Use it to confirm the connection works |
Connections made before October 18, 2026 knew these tools by earlier names (for example cellcog_create_chat and cellcog_list_employees). Those names still work for a while, so nothing breaks, but they are no longer listed; a new chat in your client picks up the names above.
A connected app can hire an AI employee and message it. It cannot delete, pause, share or reconfigure an employee directly, change your connectors, or act on your computer or browser; those stay in the web app. To change how an employee works, the app asks the employee in a message, as you would.
The five employee tools let your assistant hire an AI employee and work with the ones you already have, under the same rule as the web app: the ones you own, and the ones a teammate gave you team access to, inside the organization the connection belongs to. An employee you have team access to is instructed on its owner's behalf, so the owner's credits pay for its work; the list says which applies (access is owner or team, with the owner's name). No new permission is involved: the tools use the chat permissions you already approved, so an existing connection sees them in a new chat.
An AI employee works in numbered shifts: one workspace and memory, one conversation per shift (see the AI Employees Guide). That is why every employee tool talks in shift numbers:
An earlier shift is read-only: messages always go to the current one.
A chat started through the connector bills exactly like one you start in the web app: your credits, or your organization's shared pool if you are enrolled. An AI employee hired through the connector is yours and works on your credits, exactly as one hired on the web; a message to an AI employee bills its owner, as it does on the web. You pay for the work, not the hire; the cost depends purely on how much work you assign. Figures and the minimum balance to start a chat are on cellcog.ai/pricing.
If your balance is too low, the tool does not run and returns a structured answer instead: the ways to continue that fit your account, in order with the recommended one first, each with its link, plus one ready sentence the assistant can show you. If you have never had a subscription and a first-month offer is live, the answer carries the offer too, and clients that render inline views show it on results as a card with a countdown and a Claim offer button per plan; when the offer ends, the card becomes one line with a link to all plans. While Auto-Recharge is already topping up, the answer carries no payment link and says to try again in a few minutes. If all your parallel-chat slots are busy, the tool returns a clear error. The Account tool shows both numbers.
Files come back as links valid for 30 days, with a Download button on each, and render inline in clients that support inline views; other clients see the links. Save what you want to keep.
Some assistants run their tools in a sandbox that reaches only an allowlist of sites. The CellCog tools still work there, but the assistant itself cannot download the files it made; you can always open the links yourself. If your workspace has such a policy, ask its admin to allow cellcog.ai and storage.googleapis.com.
Connected apps are not API keys and never count toward the key limit. Deleting your CellCog account removes every grant.
The same information, with the setup steps for Cursor (one click), Claude web and Desktop, Claude Code, Codex, Muse, OpenClaw, Hermes and other agents, plus your connections, lives in the Delegate app (MCP / API in the left sidebar): cellcog.ai/delegate. Pick your agent to see every way it connects; Browse all skills lists the capabilities your agent can call by name.
"Authentication required" or the tools stop responding Your sign-in expired or was revoked. Reconnect from your client, or check Connected apps to see whether the app is still connected.
Claude does not offer the CellCog tools Start a new chat. Connectors join chats opened after you connect.
Claude asks for approval on every call That is Claude's default for a new connector. Set the tools you trust to always allow in Claude's connector settings.
The chat did not start Call the Account tool. A low balance or no free parallel-chat slot is the usual reason; the error carries the link to fix it.
An attachment was rejected Files from your computer go through Upload files (10 files, 100 MB each, kept 24 hours). Links must be public https and 25 MB or smaller; files that need a login to reach cannot be fetched.
Status is needs_input The agent asked you something. Read it with Get result and answer with Chat: send message.
I got an email saying my Claude never collected a result A chat your client started finished (or stopped with a question) and no client fetched the result within 30 minutes. Tell your client the result is ready and to fetch it with Get result; the email carries the chat id. Fetching a finished result, or sending a follow-up message, counts as seen, so a chat whose result your client collected never triggers this email.
The consent page asks me to sign up You need a CellCog account to connect. Sign up (free to try, no credit card needed); after you verify your email and accept the terms, the consent page opens again and you click Approve. The request stays open for 30 minutes. If it has expired, start the connection again from your client.