Henry Mao

The future of the internet will be dominated by tool calls, not clicks. We're building Smithery to orchestrate this new era of AI-native services for AI agents. Read our mission 👇

@ericbahn @linderps x.com/Calclavia/stat…

@linderps Where do I download that skill?

how to find a gf in sf: - dress nicely. no tech bro logos. style your hair, wear a nicely fitted tee, go to the gym - go where she goes. farmer's markets, coffee shops, dinner parties. get a sexy hobby - say hi. compliment her style, her bag, her hair. find out what she's doing there. dont bring up AI or agents lastly, don't give up. you don't find her by waiting, you find her by trying. good luck 🫶

Made a skill for this smithery.ai/skills/calclav…

@kunchenxyz This is a valid criticism. I did a simple transform to control for variability in design, otherwise it'll be too hard to compare. Ultimately, given the minor difference in performance, I think a well crafted CLI/MCP will triumph over the other.

Great analysis. However the approach of using heuristics to transform REST APIs to CLI and MCP makes this comparison not representative of the true debate. A great MCP server and agent-first CLI should not be made through a generic transformation from REST. They need to be purposefully designed, just like how much time people spend on designing a great user interface for our phones and apps. I know the article touched on the “interface design” element and dismissed it, but I think that’s part of the debate, and since the analysis is benchmarking token efficiency and success rate and trying to settle the debate with today’s models, it can’t ignore how much impact the interface design can have on the results. There’s a massive difference between a CLI / MCP that’s well optimized for agent ergonomics vs a wrapper of REST APIs, and I hope to see an analysis that truly covers that.

We're settling the MCP vs. CLI debate by benchmarking Codex and Claude Code across 3 different APIs, totaling 756 runs. We also cover skills, code mode, pretraining bias, and how MCP fits in the broader picture. The results might surprise you👇

@jainarvind It's also more performant in certain harnesses: x.com/Calclavia/stat…

MCP isn’t dead – it was just pointed at the wrong problem. On your laptop, CLIs and ad-hoc wrappers win. But at company scale, you need central auth, shared telemetry, and one integration surface for every AI host, and that’s exactly where remote MCP servers start to shine.

@RhysSullivan Skills should be reserved for user specific behavior rather than trying to replace what API vendors should maintain. x.com/i/status/20339…

skills is still not sitting right with me as a concept i think it's because companies rushed to them as the next big thing as is what happens with all ai things now everyone is their docs as skills but it's recreating all the issues (authority, up to dateness) docs solved

@ChangJonathanC We used the codex CLI default

@Calclavia > GPT 5.4 which effort level?

@brianchew We might start making flame merch if enough people want it 🤔

@Calclavia props to the design folk who did the OG design.. the mascot is really neat! look at how cute it is

@kamathematic I was this close to refactoring all our APIs to GraphQL...

We actually went into this thinking that GraphQL would come out on top due to its flexibility in request composition. Turns out MCP has its advantages, who woulda thought 🤷‍♂️

smithery.ai/blog/mcp-vs-cl… Co-author @rachpradhan

@zeeg Codebases become hard to understand when you move too fast with LLMs. What helps is trying to move more slowly despite having fast tools. Takes discipline.

im fully convinced that LLMs are not an actual net productivity boost (today) they remove the barrier to get started, but they create increasingly complex software which does not appear to be maintainable so far, in my situations, they appear to slow down long term velocity

Small feature launch of the day: You can now upload/host your Skills on Smithery instead of being limited to only GitHub URLs.

@thesophiaxu I need this!

tool i've been building: it OCRs my screen in background (like Rewind), hierarchially-summarizes into a timeline, then makes it available via a local api then i just point claude code at it and ask it to identify my inefficient workflows, and it found a bunch

Annual plans for AI subscriptions are now a nonstarter. You have to be ready to hop to the best tool of the month.

Is it just me, or are CAPTCHAs hard enough now that I'm starting to fail more?

@oscrhong As you're executing a long running task and you notice it goes in the wrong direction, you can point it in the right way!

@Calclavia any favourite uses?

Steering on Codex is an underrated feature

@stuffyokodraws @openclaw Which APIs are you using?

You open @openclaw today thinking you will be vibe automating and vibe coding 10 hrs later you are just vibe copy pasting API keys and creating 11th account for Openclaw Can someone solve this integration problem for agents

@karpathy Agents today are productive, not yet creative

I had the same thought so I've been playing with it in nanochat. E.g. here's 8 agents (4 claude, 4 codex), with 1 GPU each running nanochat experiments (trying to delete logit softcap without regression). The TLDR is that it doesn't work and it's a mess... but it's still very pretty to look at :) I tried a few setups: 8 independent solo researchers, 1 chief scientist giving work to 8 junior researchers, etc. Each research program is a git branch, each scientist forks it into a feature branch, git worktrees for isolation, simple files for comms, skip Docker/VMs for simplicity atm (I find that instructions are enough to prevent interference). Research org runs in tmux window grids of interactive sessions (like Teams) so that it's pretty to look at, see their individual work, and "take over" if needed, i.e. no -p. But ok the reason it doesn't work so far is that the agents' ideas are just pretty bad out of the box, even at highest intelligence. They don't think carefully though experiment design, they run a bit non-sensical variations, they don't create strong baselines and ablate things properly, they don't carefully control for runtime or flops. (just as an example, an agent yesterday "discovered" that increasing the hidden size of the network improves the validation loss, which is a totally spurious result given that a bigger network will have a lower validation loss in the infinite data regime, but then it also trains for a lot longer, it's not clear why I had to come in to point that out). They are very good at implementing any given well-scoped and described idea but they don't creatively generate them. But the goal is that you are now programming an organization (e.g. a "research org") and its individual agents, so the "source code" is the collection of prompts, skills, tools, etc. and processes that make it up. E.g. a daily standup in the morning is now part of the "org code". And optimizing nanochat pretraining is just one of the many tasks (almost like an eval). Then - given an arbitrary task, how quickly does your research org generate progress on it?