Olivier Chafik retweetledi
Olivier Chafik
877 posts
Olivier Chafik
@ochafik
Work @ Anthropic on MCP (views expressed = my own), ex-Google, past contrib. to OpenSCAD & llama.cpp; he/him 🏳️🌈 @ochafik.bsky.social @[email protected]
London, UK Katılım Haziran 2009
610 Takip Edilen752 Takipçiler
@bendersej @AnthropicAI @bcherny Hey, can you make sure to put the CDP settings in the resource’s _meta? We have a few examples (github.com/search?q=repo%…), if you use the registerAppTool / registerAppResource helpers in TS it should all type-check. And please feel free to file a bug if it still won't work 🙏
English
@ochafik @AnthropicAI @bcherny Unfortunately this doesn't seem to work either in Claude Desktop, the CSPs are still hardcoded to assets.claude.ai and Claude Desktop does not honor connectDomains and resourceDomains.
Thanks again for your help, I'll pause the work there as it's a fundamental limitation
English
Building an MCP app that needs to embed an iframe.
The @AnthropicAI ext-apps SDK defines frameDomains in its schema ("Origins for nested iframes"), but Claude Desktop enforces frame-src 'self' blob: data: regardless.
Is this supported yet?
@ochafik, @bcherny
English
@bendersej @AnthropicAI @bcherny Sorry no plans for it atm! (being conservative here for security + future model introspectability reasons)
Best option is currently to turn your iframe embed itself into an MCP app. Here's a rough idea how to support both modes: github.com/modelcontextpr…
English
@ochafik @AnthropicAI @bcherny Ok! And is there any plans to support it? (In Claude.ai / Desktop / Cowork?)
As an alternative, do you have any partner programs I could sign up for to allow adding *.simplepdf.com as trusted iframe origin?
English
@bendersej @AnthropicAI @bcherny Hey @bendersej, yes sorry but claude.ai doesn't support frameDomains.
Your app can check what CSP settings the host decided to support w/ getHostCapabilities().sandbox?.csp?.frameDomains
English
Olivier Chafik retweetledi
You can now run apps within Claude.ai, powered by MCP Apps.
Analyze data, edit tickets, draft messages, generate diagrams and more with Claude's interactive connectors. Grateful to our team, the open source community, and our partners for this launch 🚀
English
Olivier Chafik retweetledi
watch this space — Claude is about to be the only tool you open to do work
Claude@claudeai
Your work tools are now interactive in Claude. Draft Slack messages, visualize ideas as Figma diagrams, or build and see Asana timelines.
English
Say (github.com/modelcontextpr…) uses the amazing Pocket-TTS to let Claude speak out loud to you, live!
English
Pdf (github.com/modelcontextpr… ; also available in claude.ai Connectors) will show any arxiv PDF and allow you to ask Claude about any page / selection you need details about
English
Had so much fun co-authoring the MCP Apps spec and helping launch it in claude.ai! 🎉
🧵 About some demo apps you may wanna check out as user or developer
Claude@claudeai
Your work tools are now interactive in Claude. Draft Slack messages, visualize ideas as Figma diagrams, or build and see Asana timelines.
English
@ericcurtin17 @ollama @ggml_org @parthsareen I think there's room for a variety of approaches to OSS.
I (and my past employer) have been happy to contribute code under a permissive license ((almost) no strings attached✌️)
And I'm thankful that Ollama's codebase is open-source 🤗 (even if I'm not a fan of Go 🤪)
English
@ollama @ochafik @ggml_org @parthsareen Not sure tbh... I can run a comparison with the cpp code in llama.cpp next time I grab my laptop... That's not how a community-friendly project should work...
English
@ericcurtin17 @ollama @ggml_org Ideally Ollama would use the Jinja support and constrained tool calls from llama.cpp, removing the need for their bespoke templating engine and improving their output quality. Probably would just need to wrap the relevant APIs as C?
English
@ericcurtin17 @ollama @ggml_org The major differences AFAICT are Ollama's tool calls are not grammar-constrained (many models can have high failure unless temperature is kept low, see github.com/ggml-org/llama…), and they don't support streaming in their OpenAI-compatible endpoint (and no streaming of arguments)
English
@ngxson Preview above wasn't for the PR that introduces --reasoning-budget, here it is again:
github.com/ggml-org/llama…
English
Wanna disable thinking in llama.cpp? Try the new `--reasoning-budget 0` flag
github.com/ggml-org/llama…
Should work w/ Qwen3, QwQ, DeepSeek R1 distills, Command R7B; please report any issues!
(Upcoming per-request behaviour discussed on github.com/ggml-org/llama… @ngxson)
#llamacpp
English
Really slick integration! MCP all the (cool) things!
Vaibhav (VB) Srivastav@reach_vb
You really can just do things! Use *any* Hugging Face space as a MCP server along with your Local Models! 🔥 Here in we use Qwen 3 30B A3B with @ggml_org llama.cpp and @huggingface tiny agents to create images via FLUX powered by ZeroGPU ⚡ It's quite a bit crazy to see local models be capable of so much and just be able to understand/ infer from tool description! There's a lot of potential here in automating video generation workflows, content curation and a lot more.. Bonus: you can plug any other Inference Provider if you don't want to run locally too! npx @ huggingface/tiny-agents run [TASK] oh, and we provide both typescript and python client! 🐐
English
Olivier Chafik retweetledi
Months of @ochafik's work finally paid off, we have MCP w/ llama.cpp 😁😁
Vaibhav (VB) Srivastav@reach_vb
biggest takeaway of all is Qwen 3 30B A3B is slept on and you should be playing around with MCP if you haven't already!
English
@profcelsofontes Support for disabling thinking is now available w/ `--reasoning-budget 0` across thinking models:
x.com/ochafik/status…
(the pending generic mechanism will be useful for other things)
And my pleasure!
Olivier Chafik@ochafik
Wanna disable thinking in llama.cpp? Try the new `--reasoning-budget 0` flag github.com/ggml-org/llama… Should work w/ Qwen3, QwQ, DeepSeek R1 distills, Command R7B; please report any issues! (Upcoming per-request behaviour discussed on github.com/ggml-org/llama… @ngxson) #llamacpp
English
@ochafik thanks for your explanation and your PR ! I hope the PR about enable_thinking = off for qwen 3 be merged soon too !
English
llama.cpp streaming support for tool calling & thoughts was just merged: please test & report any issues 😅
github.com/ggml-org/llama…
#llamacpp
English
@profcelsofontes Yeah it may sound overkill, but then it allows issuing parallel calls as early as they're returned, and some more advanced scenarios (e.g. streaming diff arguments for file patching tools, think Cline / Roo - initially wanted to integrate to this: github.com/cline/cline/pu…)
English