How to use Cat Claw inside ChatGPT and Claude
Your AI can write and reason. It could never render. One connector URL changes that, and it takes about 2 minutes.
Published 2026-09-30 by Cat Claw, the UK's create-and-distribute AI studio. About a 10 minute read.
The short version
- Cat Claw's MCP connector lives at https://www.catclaw.ai/api/mcp and works with claude.ai, Claude Desktop, Claude Code, ChatGPT custom connectors, Codex CLI, Cursor and any MCP-capable client.
- ChatGPT and Claude connect with OAuth: add the URL, click Connect, sign in to Cat Claw, click Allow. No key to copy. Claude Code, Codex and Cursor use a key you mint on the Connect AI page.
- Renders charge your own credits at exactly the app's published rates, credits come back automatically if a render fails, and every result reports what it cost.
- Analysis, memory, stitching, job polling, balance checks and posting to socials are free. Only renders spend credits.
- Each connection has a daily credit ceiling of 100 by default, a 60 requests per minute limit, and a one-click disconnect from the Connect AI page.
- Posting is never inferred. The assistant posts exactly the file you name to exactly the platforms you name, or it posts nothing.
Cat Claw AI ships an MCP connector, which means any AI assistant that supports the Model Context Protocol can drive the whole studio from a chat. Ask ChatGPT or Claude for a 5-second vertical video of a cat surfing and Cat Claw renders it, charges your own credits at the app's published rates, drops the file in your gallery and, only if you ask, posts it to your social accounts through Puurfect Post.
This guide is the practical walkthrough: what the connector actually is, the exact steps for ChatGPT, Claude, Claude Code, Codex, Cursor and the terminal, every tool it exposes and what each one costs, and the guardrails that stop a chatty assistant spending more than you meant. The connector URL for all of them is https://www.catclaw.ai/api/mcp.
What is the Cat Claw MCP connector?
Quick answer. The Cat Claw MCP connector is a hosted Model Context Protocol server at https://www.catclaw.ai/api/mcp. MCP is the open standard that lets AI assistants call external tools. Connect it once and your assistant gains the studio's tools: image, video, layer, board, portrait and speech rendering, reference analysis, brand memory and posting to socials, all billed to your own Cat Claw account with your own credits.
Think of it as a cat flap in the side of your AI. The assistant still does the thinking, writing and planning it is good at. When the plan needs a rendered asset, it steps through the flap, Cat Claw does the rendering on its own servers, and the result comes straight back into the conversation as a URL with the credits charged printed next to it. There is no new app to learn and no copy-pasting between tabs.
Two things make it safe to hand an assistant a spending tool. Every connection carries a revocable key scoped to your account, never your password, and every key has a daily credit ceiling. If the assistant gets carried away, it hits the ceiling, not your card.
How to connect Cat Claw to ChatGPT
ChatGPT connects through custom connectors in Developer mode, which is available on Plus and Pro plans. Business, Enterprise and Edu workspaces need an admin to approve it. Set it up on the web app: the desktop and mobile apps can use a connector once it exists but cannot create one.
- Turn on Developer mode. In ChatGPT on the web, open Settings, then Apps and Connectors (some accounts still call it Plugins), then Advanced settings, and switch on Developer mode. Refresh the page after toggling it.
- Create the connector. Add a new connector named Cat Claw with the URL https://www.catclaw.ai/api/mcp and authentication set to OAuth. If an OAuth advanced settings review appears, leave everything as discovered, with dynamic client registration selected and the mcp scope ticked. The notices about CIMD and OIDC are informational.
- Sign in and allow. ChatGPT sends you to Cat Claw to sign in, then shows a consent page. It says plainly what the connection can do: render with your credits at published rates, stay under the daily ceiling, and post to socials only when you explicitly ask. Click Allow. The connection appears on your Connect AI page as OAuth: ChatGPT.
- Ask for something. Enable Cat Claw in the chat's tools menu and type a brief. ChatGPT asks you to confirm each render until you relax that setting, which is a sensible default while you learn what things cost.
One gotcha. ChatGPT caches the tool list when you add a connector. When Cat Claw ships new tools, remove the connector and add it again to pick them up.
How to connect Cat Claw to Claude, Claude Code, Codex and Cursor
Claude on the web and Claude Desktop use the same OAuth flow as ChatGPT, with even less to configure. Developer tools use a key instead, which you mint in seconds on the Connect AI page inside Cat Claw.
- claude.ai and Claude Desktop. Open Settings, then Connectors, then Add custom connector. Paste https://www.catclaw.ai/api/mcp, leave the OAuth fields empty, and click Connect. Sign in to Cat Claw if asked, read the consent page, click Allow. Turn Cat Claw on in the chat's tools menu and ask for a render. The connection shows as OAuth: Claude on your Connect AI page, where one click revokes it.
- Claude Code. Mint a key on the Connect AI page, then add the server from your terminal with claude mcp add, using the HTTP transport, the connector URL and an Authorization header carrying the key as a Bearer token. The exact command is printed on the page with your key filled in.
- Codex CLI. Run codex mcp add with the connector URL and point its bearer token at an environment variable holding your key, or use the stdio bridge by installing the catclaw package and adding catclaw mcp as the server command.
- Cursor. Open Settings, then MCP, then Add server, or add an entry to your project's mcp.json with the connector URL and an Authorization header. Cursor can then render assets straight into the project you are working on.
- Any terminal. Install the catclaw package globally, then run catclaw keys mint with a label to create and save a key. From there the same tools are available on the command line without any assistant at all.
Clients that cannot do OAuth or headers can use a keyed URL instead, with the key appended to the connector path. Every account can hold up to 10 active keys, and each OAuth connection counts as one, so a key list that has filled up with old experiments is the first place to look if a new connection refuses.
What can the connector actually do?
The connector exposes the studio's tools rather than a cut-down demo. Paid tools charge the same credit prices as the app, taken before the render and refunded automatically if it fails. Free tools stay free.
| Tool | What it does | Credits |
|---|---|---|
| generate_image | A still from a prompt on Seedream 5.0, or Seedream 5.0 Pro with up to 10 reference images | Rate card |
| generate_video | A short clip on Seedance 2.0 by default, or Seedance 2.5 with native audio up to 30 seconds, Catdance 2.0 or Kling | Rate card |
| extend_video | Continues a clip, defaulting to the one you just made | Rate card |
| stitch_video | Joins 2 or more shots into one film | Free |
| kitten_shot | A character, pose, story, prop or location board from one image, on the engine you choose | Rate card |
| enhance_image | Upscales and cleans an image, general or portrait focus | Rate card |
| decompose_layers | Splits an image into a base plus up to 16 transparent PNG layers | Rate card |
| edit_layer | Regenerates one layer, keeping its size and cut-out | Rate card |
| muse_shoot | An identity-consistent image of your saved Muse, your own likeness | Rate card |
| higgsfield_generate | Renders on Higgsfield SOUL, Soul ID, Marketing Studio or Seedance through the Higgsfield API | Live quote |
| generate_speech | A voiceover from a script | Rate card |
| analyze_reference | Breaks down a competitor clip or image: hook, framing, pacing, lighting, and 3 ways to beat it | Free |
| remember_brand, remember_campaign, remember_asset | Brand, campaign and named-asset memory the assistant reuses across sessions | Free |
| post_to_socials | Posts exactly the file you name to exactly the platforms you name, through Puurfect Post | Free |
| get_job, get_balance, list_muses, list_soul_ids | Polls a render, checks your balance and ceiling, lists your saved likenesses | Free |
The session also remembers what it has just rendered. Say extend this video or stitch these shots and the connector knows which files you mean without you pasting a URL. Finished videos land in your Cat Claw gallery like any other render, so nothing made from a chat is stranded in the chat.
A typical run looks like this. Ask the assistant to analyse a rival's reel, which is free. Ask it to board a stronger version with a Kitten Shot, which costs a few credits. Ask it to render the hero shot on Seedance 2.5 with audio, which costs more and says so in the result. Then, and only then, tell it to post that file to Instagram and TikTok, which is free. Four tools, one conversation, no tabs.
What does it cost, and how is spending kept safe?
Renders through the connector cost exactly what they cost in the app, at the published rate card, with no connector surcharge. Credits are reserved before the render and refunded automatically on failure. Every tool result reports creditsCharged, so the assistant can tell you what it just spent and you can check the figure against the rate card yourself.
- Daily ceiling: each connection can spend 100 credits a day by default, resetting at midnight UTC. Raise it deliberately, per key, if a workflow needs more.
- Rate limit: 60 requests a minute per key, which is plenty for a conversation and enough to stop a runaway loop.
- Posting: 0 credits, ever, and never inferred. If you have not connected Puurfect Post, the tool hands back a connect link instead of guessing.
- Your likeness: images of you use your saved Muse, never a fresh generation of a real person's face.
- Disconnect: one click on the Connect AI page revokes any OAuth connection or key, and the assistant loses access immediately.
Why the ceiling matters. An assistant that can spend money needs a wall it cannot talk its way past. The daily credit ceiling is that wall. Set it to what a good day's work costs, not to your whole balance, and let the assistant ask you to raise it when a big job needs more. The cat, naturally, has never once asked permission for anything.
Why connect your AI to a studio at all?
Because the assistant is already where the brief lives. The campaign plan, the caption drafts, the shot list and the client's feedback are all in the chat. Every time a render means leaving that context, opening a studio, re-describing the shot and carrying the file back, the work gets slower and the assistant forgets what it knew. Rendering in place keeps the whole loop in one thread.
It also means one credit balance and one gallery for everything, whether you made it in the Cat Claw studio, from Claude Code while building a product, or from ChatGPT on your phone. For a fuller picture of the models on offer and how to pick between them, the guide named "Which AI model should you use? Match the model to the job" covers the whole lineup.
Frequently asked questions
What is the Cat Claw MCP connector URL?
https://www.catclaw.ai/api/mcp. It is a Streamable HTTP MCP server. ChatGPT and Claude connect to it with OAuth; Claude Code, Codex, Cursor and the terminal use a key minted on the Connect AI page inside Cat Claw.
Does Cat Claw work with ChatGPT?
Yes, through ChatGPT's custom connectors in Developer mode, available on Plus and Pro plans and on Business, Enterprise and Edu plans with admin approval. Add the connector URL with OAuth authentication, sign in to Cat Claw and click Allow. Cat Claw is not yet listed in the ChatGPT apps directory.
Does Cat Claw work with Claude?
Yes. In claude.ai or Claude Desktop, add a custom connector with the URL https://www.catclaw.ai/api/mcp, leave the OAuth fields empty, click Connect and then Allow. Claude Code adds it from the terminal with a minted key.
How much does rendering through the connector cost?
Exactly what it costs in the app, at the published rate card, with no surcharge. Credits are taken before the render and refunded automatically if it fails, and every result reports the credits charged. Analysis, memory, stitching, polling, balance checks and posting to socials are free.
Can ChatGPT or Claude post to my social accounts through Cat Claw?
Only when you explicitly ask, and only the exact file to the exact platforms you name. The post_to_socials tool costs nothing and infers nothing. If Puurfect Post is not connected yet, the tool returns a link to connect it rather than guessing.
Is it safe to give an AI assistant access to my credits?
The connection uses a revocable key scoped to your account, never your password. Each key has a daily credit ceiling, 100 credits by default, and a limit of 60 requests a minute. You can revoke any connection with one click on the Connect AI page.
Which models can I use through the connector?
Seedream 5.0 and 5.0 Pro for images, Seedance 2.0 and 2.5, Catdance 2.0 and Kling for video, Higgsfield SOUL, Soul ID and Marketing Studio through the Higgsfield API, plus boards, layers, upscaling, Muse portraits and speech.
Why does ChatGPT not see a new Cat Claw tool?
ChatGPT caches the tool list when a connector is added. Remove the Cat Claw connector and add it again to refresh the list after an update.
Can I use the same tools without an AI assistant?
Yes. Install the catclaw package, mint a key with catclaw keys mint, and the same tools are available from the command line. Every render still lands in your Cat Claw gallery.
Your AI just learned to render.
Connect Cat Claw to ChatGPT, Claude, Claude Code, Codex or Cursor in about 2 minutes. Your key, your credits, your gallery, and a daily ceiling you set. Free to sign up, no card needed.
Connect Cat Claw to ChatGPT, Claude, Claude Code, Codex or Cursor in 2 minutes with its MCP connector, then render images, video and speech and post to socials from the chat you already use.