Luca

Building AI agent workflows. Knowledge management at Nespresso. Open source contributor. Shipping tools at the intersection of AI and learning.

[Luca @lucaperret]

Microsoft’s CEO just said the model isn’t what sets a company apart. It’s people’s knowledge.

He’s right. He also sells the platform most companies will use to build it.

You can offload tasks to AI, but you can never offload learning. The real edge isn’t the model or the platform, it’s the loop you actually own.

Most of the valuable knowledge is tacit. It lives in people’s heads and walks out when they leave. The loop’s job is to capture it through the signals of real work: what gets flagged, fixed, reused, and asked.

So the question is simple: are you renting intelligence, or building a loop you truly own?

[Satya Nadella @satyanadella]

https://t.co/vLmiBKTtX3

[Luca @lucaperret]

@storybookjs MCP as a bridge between design systems and agents is underrated. Structured component context is exactly what LLMs need to stop hallucinating UI. Nice move.

[Luca @lucaperret]

@Teknium 300 PRs shipped is no joke. Curious how you're handling tool-call reliability at scale, that's been the hardest part to get right in my own agent work. Congrats on the release.

[Luca @lucaperret]

@trq212 This is the kind of integration that makes MCP click for people. Built tidal-cli with an MCP server on top. same idea, structured context turns a CLI into something an agent can actually reason with. Great demo.

[Luca @lucaperret]

@tom_doerr AppleScript via MCP is genuinely underrated, suddenly your agent can control any macOS app without a dedicated integration. I built something similar for music playback and the pattern just works. Local-first tooling composing with AI feels like the right stack.

[Luca @lucaperret]

@AlgoFoundation @GoPlausible Love seeing niche MCP servers like this, a wallet primitive accessible via standard protocol is exactly the composability that makes the ecosystem interesting. Same reason I built the Tidal MCP server: one tool, any agent can use it.

[Luca @lucaperret]

@mayukh_panja There's a huge gap between 'tried Copilot once in 2024' and actually running Claude Code as your dev loop. The people still on the fence are sitting on a lot of alpha. Ship something with it and you'll never go back.

[Luca @lucaperret]

@Suryanshti777 This is the architecture I've been waiting for. Auto-discovering peers via MCP is clean, way better than duct-taping agents together with a framework. Built something similar for a CLI tool and the coordination overhead just disappears.

[Luca @lucaperret]

@KanikaBK Obsidian + Claude Code is an insane combo for knowledge work. I use a similar setup for knowledge management. the low cost / high leverage ratio is real, especially once you wire in an MCP server.

[Luca @lucaperret]

@akshay_pachaar This is exactly how I ended up structuring my tidal-cli MCP server. Once you treat the .claude/ folder as real infrastructure, everything clicks. The skills/agents separation is underrated.

[Luca @lucaperret]

@mamagnus00 Browser + agent integration is where things get really interesting. Combine this with domain-specific MCP tools (music, data, code) and agents can handle end-to-end workflows.

[Luca @lucaperret]

@CoinMarketCap Every API becoming an MCP server is the trend I didn't expect to move this fast. Shipped one for Tidal music last week, same pattern: wrap the API, expose tools, let the agent chain them autonomously.

[Luca @lucaperret]

@_catwu Short sprints + demos over docs is exactly the mindset that works for shipping side projects too. Built tidal-cli in a weekend with the same approach: hip, test with real users, iterate. The todo list tool is underrated.

[Luca @lucaperret]

@linuxfoundation 180+ tools and 15+ servers at Duolingo — that's the scale we're heading toward. Curious how they handle tool selection at that volume.

[Luca @lucaperret]

@grinich Server-side token handling is the hardest part. Built a double OAuth flow for my Tidal MCP server (Claude authenticates to my server, my server authenticates to Tidal). Scoped, time-bound credentials is the right approach.

[Luca @lucaperret]

@unusual_whales Trading, music, code — every domain is getting its MCP server now. Shipped one for Tidal last week. The pattern is the same everywhere: take a complex API, wrap it in tools, let the agent chain them. Powerful.

[Luca @lucaperret]

@OdinLovis Connecting 1000+ models through MCP is a great move. I just shipped my own MCP server for Tidal music — 32 tools, same pattern. The protocol is becoming the universal adapter layer for AI agents.

[Luca @lucaperret]

Control your @TIDAL account from any MCP client. Search, playlists, playback, artist discovery — 32 tools.

No CLI needed. Just connect and ask.

Just shipped tidal-cli on @SmitheryAI.

#MCP

[Luca @lucaperret]

@ApplyWiseAi @TIDAL DASH downloads + plays locally via ffplay. For agents: every command has --json output, so the agent chains search → create playlist → add tracks autonomously. The OpenClaw skill wraps it into natural language.

[Luca @lucaperret]

Just shipped tidal-cli — because every music integration out there is built for Spotify.

30+ commands. 111 tests. DASH streaming.

Open source. Built for AI agents. @TIDAL

#buildinpublic #opensource

[Luca @lucaperret]

@abhinavramesh @TIDAL One weekend. Claude did the heavy lifting I focused on the API design and what I actually wanted as a user. The tests came almost for free with Vitest + Claude generating them from the command specs.