Allow users to connect their own Claude or ChatGPT subscriptions to power Framer AI
C
Christian Brescia
I’m loving the new agent-building feature and was excited to use it to resolve a long-standing issue on my homepage. I spent considerable time working with the agent, providing detailed prompts, screenshots, GIFs, and other context, but ultimately exhausted a significant portion of my monthly AI allotment without resolving the problem. In fact, I’m likely going to delete the branch it created because the proposed changes weren’t successful.
I understand that AI compute has real costs and that refunding usage after changes have been generated may not be practical. However, I think there is an even better long-term solution.
Please consider allowing users to authenticate Framer with their own Claude and/or ChatGPT accounts (or API keys) so Framer’s AI features can leverage subscriptions they already pay for. Many of us already have premium plans with these providers and would prefer to use those resources while continuing to work inside Framer’s excellent native development environment.
This would:
* Reduce friction for power users who already pay for external AI services.
* Eliminate concerns about quickly exhausting Framer’s AI allotment on complex debugging tasks.
* Encourage deeper adoption of Framer’s AI tools without forcing users to choose between platforms.
* Keep developers in Framer instead of pushing them to external tools like Claude Code for longer sessions.
I think this would be a significant quality-of-life improvement for advanced users and teams who rely heavily on AI-assisted development.
Log In
Autopilot
Merged in a post:
Allow users to connect their own Claude or ChatGPT subscriptions to power Framer AI
H
Holly Kapp
can you make it so that I can use my own subscription in the future because I just hate having to pay for extra tokens on months I need to make website updates!!!
S
Somesh K S
Connecting external agents is so shity!
Session instability — Multiple exec calls crashed with "Cannot read properties of null" and the cryptic Assertion failed: !(handle->flags & UV_HANDLE_CLOSING) error. This suggests the session connection was dropping or the node tree wasn't loading properly.
Wrong node IDs — I was using node IDs from the earlier conversation (like R213M_bzz for the case study section) without verifying they still exist or are accessible in the current project state. Just because we referenced them before doesn't mean they're still there or reachable.
Depth parameter confusion — I kept trying different depth values (0, 1, 2, 3) without understanding whether the projects page actually has children to serialize. The page might be dynamically rendered or use a different structure than expected.
Assuming page structure — I assumed the projects detail page (y0Uy_iYD7) would have a standard frame/variant hierarchy, but it came back with 0 children. The structure might be completely different (maybe it's using CMS bindings, code components, or a different layout pattern).
Not checking Framer docs first — Per the skill instructions, I should have run npx @framer/agent@latest docs to verify the exact API signatures before writing code. Instead I just guessed at method parameters.
Too many small calls — Each failed call burned time and context. I should have either (a) loaded the project once, stored everything in state, then analyzed it, or (b) used framer.agent.readProject to get a bulk snapshot of the page structure upfront.
Requesting permission without context — The last attempt got rejected because I hadn't explained what I was doing first, and the tool use looked like it might be exploratory rather than purposeful.
S
Stefano Ferragina
Completely agree — thanks for articulating this so clearly.
One additional point: the current credit system feels quite opaque. Allowing API integrations would give users full visibility and control over costs, which is especially important for advanced use cases.
R
Rooz !
I was about to write a new thread but I am doubling down on this. I had a long chat with the framer team. The current agent is super clunky and it is inefficient in understanding the task in hand. I had the exact same experience with the agent - drifting like crazy.
Kudos on bringin the agent to the platform. It is not for the prime time yet. I am happy to experiment with my Claude subscription but paying extra and get frustrated - not so much :D