Hi @justmert , the event takes place on Thursday at 3pm CET in the ICP Discord’s voice channel.
This is the voice channel: ICP
This is the event (which also shows the time in your timezone): ICP
Please let me know if you have any other questions. See you then 
Hi everyone, thank you for today’s call (2025.08.07). Special thanks to @justmert for the demo! This is the generated summary (short version, please find the long version here ): The ICP Agent Kit, presented by Mert in the DeAI Working Group, is a TypeScript library designed to simplify interaction with Internet Computer (IC) canisters using either direct API calls or natural language prompts processed via OpenAI. It features a modular plugin system (canister, token, identity, cycles) and supports rapid agent development, ideal for educational and developer onboarding use cases. While not an MCP or A2A-ready tool yet, it lays the groundwork for such integrations. Security-wise, prompt clarity and validation are key, especially before mainnet use. Looking ahead, the project aims to evolve into a Language Model-Canister Protocol (LMCP) to enable AI-assisted development in environments like Cursor. Broader discussion emphasized the trend toward small language models for edge use, WebAssembly deployment, and IC’s strategic potential to differentiate through tooling and agent-centric infrastructure.
Hey @patnorris, thanks for sharing the summary from the recent DeAI Working Group call! Love to see the momentum around AI-assisted development tools for ICP.
I wanted to mention that I will be implementing dev-focused MCP server named ICPilot that directly addresses the need for MCP integration that was discussed. ICPilot enables natural language interaction with the ICP developer tool ecosystem through conversational commands in any MCP-supported environment like Claude Code, Windsurf, Cursor, or Gemini CLI.
No need to switch between multiple tools, no more hunting through documentation, and no more complex CLI commands or configurations. Just tell ICPilot what you want to do in natural language while developing your ICP application in Cursor or any MCP compatible dev tool , and it will handle the rest for you, for developers.
I have already sent the project proposal for ICPilot, which provides full dfx CLI integration through MCP along with comprehensive tool integrations including Mops, Quill, and didc. The project also features RAG-powered ICP documentation search and supports all major MCP clients, directly addressing the MCP needs discussed in the working group call.
—
You can see the demo and details at: GitHub - tolgayayci/icpilot: Your AI Assistant for Internet Computer Development
You say: “Create a new dfx project called hello_world”
What happens: Your AI assistant will create a new Internet Computer project with all the necessary files.
It seems like perfect timing that the community expressed interest in MCP solutions right as I submitted the proposal. I’d love to contribute to filling this gap in our developer tooling ecosystem and would welcome any feedback. @marc0olo @cryptoschindler @ddave
P.S: For who don’t know me I am also developer of DFX Dashboard, which helps you to interact with dfx cli via specialized GUI.
GitHub: GitHub - tolgayayci/dfx-dashboard: DFX Dashboard is a user-friendly GUI built on top of the dfx CLI, simplifying canister management, network configuration, and project workflows on the Internet Computer.
Docs: DFX Dashboard Documentation
Hi @tolga , thanks for reaching out. Your work on MCP sounds interesting, and this week’s call fits really well as @baolongt will present their latest work with MCP on ICP
Hi everyone, thank you for today’s call (2025.08.14). Specials thanks to @kyle , @baolongt and @yrgg for presenting their latest work! This is the generated summary (short version, please find the long version here ): The call focused on recent progress in building MCP (Model Context Protocol) on the Internet Computer, including OAuth 2.0 support in the Rust SDK for secure, multi-user authentication via Internet Identity. This enables use cases like AI-assisted crypto wallets and multi-user wallet management. Discussions covered deploying MCP servers on-chain to improve trust through WASM hash verification, creating a trustless MCP Server Marketplace, and potential integration with existing ICP canisters using auto-generated MCP interfaces. Additional discussions included running AI models as WASM modules in canisters, and unlocking decentralized AI workflows with ICP’s large memory capacity.
Hi everyone, join us for our next hardware- and infrastructure-focused session (Accelerated Infrastructure for AI on ICP) this Thursday! @icarus will lead the call Looking forward to it, and see you then.
Hi everyone, are there any items you would like to see on this week’s agenda? Or is there anything you would like to share/present during the call?
Hi
@patnorris Would love to discuss regarding
and further regarding ICP LLM privacy.
Hey @patnorris ,
We released the lib and some numbers today. It’s 7x faster than any other verifiable AI (zkML w/ ONNX) packages, in-spite of only using CPU at the moment. I would not be surprised if the ICME team hits 30x this year — taking verifiable AI times down from minutes to seconds.
Hi @MalithHatananchchige , Hi @apotheosis , thank you, these sound like great topics to discuss 
looking forward to it!
Hi @KairosOPs , thanks, yep, we keep summaries of all calls in this repo: DeAIWorkingGroupInternetComputer/WorkingGroupMeetings at main · DeAIWorkingGroupInternetComputer/DeAIWorkingGroupInternetComputer · GitHub
Hi everyone, thank you for today’s call (2025.08.28). Special thanks to @MalithHatananchchige and @icarus for sharing their latest insights and research! This is the generated summary (short version, please find the long version here ): The call covered three main threads: first, experiments with alternative AI hardware showed current custom setups struggle to match NVIDIA’s performance, though FPGA-based startups like Positron AI offer interesting reprogrammable options; second, privacy and trust remain major risks in centralized AI (e.g., log retention, model substitution), with TEEs and IC-local execution seen as credible paths to provable confidentiality; third, discussion turned to architecture, where the group explored integrating confidential GPU-backed VMs directly into ICP subnets via consensus-driven deployment and VSOC connections, though the key open question remains whether the added security and decentralization benefits justify the cost and complexity compared to simply using external APIs.
