scott

@rozzabuilds I’m sure you already have but make sure since you are using thier tool that don’t get rights or access to other projects or work you do with it.

I CANCELLED MY CLAUDE SUBSCRIPTION It’s been a long time coming tbh. I needed to prepare for the next stage of my journey. I didn’t want to pay for something that jut wasn’t necessary any longer. And yes, I love Claude. This isn’t something I anticipated happening any time soon. … but as a client offered to pay for Claude going forward on my behalf, who am I to reject that kind of offer 🤷‍♂️

@capajj @aidenybai It does not need to be this way! github.com/srhall2314/thi…

@aidenybai fuck the MCP, CLI FTW

the current state of MCP

@graphtheory @streuselberg Fully agree.

@theScottrHall @streuselberg Tool definition overhead is very much not negligible at current context window sizes with the number of tools needed by a general purpose production grade agent

MCP is already legacy. It’s amazing tbh. Even in their own training materials, Anthropic recommends CLI + SKILL.md.

@streuselberg @graphtheory I think you’re right this is very use-case dependent.

@graphtheory Do you think that's better than MCP in your opinion? It could largely be a result of the usage context but in an editor I feel like mcp usually performs better and more deterministically but I haven't really done a close comparison or anything

@tkSoftwareCat Yes, it does. But you get the befits of MCP but reduce token usage by having MCP point directly to the API instead controlling the communication. See github.com/srhall2314/thi…

ふーむCLIはMCPと比較してトークン消費量の優位性があるのか qiita.com/toarusyakaijin…

@saen_dev @tpschmidt_ I agree. I use usejoashua.com it also directs the agent to the API directly instead running through the MCP connection, so less token bloat.

@tpschmidt_ Configuring the same MCP server four times across four clients was genuinely the most annoying part of the whole ecosystem. Install once and works everywhere is how it should have been from day one.

Docker just solved my most annoying problem with AI tool integrations: Configuring the same MCP server separately for Claude Desktop, Claude Code, Cursor, and whatever else you use? Yeah, that's over via the MCP Catalog Just install once → works everywhere

@keeran @kdaigle Why CLI? Just give your agent the API. If it needs key and context you can do that in just a .md file.

@kdaigle how many times have you felt the ‘gah im just gonna wrap this up in a cli’ urge? since the -codex models came out ive barely used MCP, preferring to build simple cli (skills) tooling for everything.

While I don’t tend to get a bunch of midweek coding time, Saturday afternoons are all coding and using everyone’s tools. Polished up my flight search tool. Let 5.5 crush an insidious bug I had in my household management app. Still feels like MCP setup is stuck in 2005.

@plainlogic24 @samgrows @sahil_builds API all the way. I mean CLI and MCP are both just wrappers for the true primitive the API. Now you can use thin MCP to get out of the way if you need the discovery of MCP. github.com/srhall2314/thi…

@samgrows @sahil_builds No mcp server please create a cli or api it will be easy to integrate with skills otherwise mcp server will be running all the time even if i want it or not

BREAKING: Claude can now watch videos for you! You can plug in a YouTube, Tiktok or Instagram URL and let Claude break it down for you Live now inside of Algrow

@tkSoftwareCat Yes, but you can also really reduce MCPs token usage by having it only connect your agent to the API not mediate the connection. github.com/srhall2314/thi…

Created a repo to reference thin-mcp. github.com/srhall2314/thi…

Correct. MCP and CLI are simply wrappers for APIs. Agents can go straight to the API if provided with three things: api docs, auth structure and context. So a thin MCP can deliver that and get out of the way reducing tokens and increasing speed . UseJoshua.com does this.

Sometimes I think even Naval doesn't know what he is talking about. What do you mean AI will replace APIs?

Thin MCP The best use of MCP is to make an introduction not to carry the conversation. Any agent or chat bot understands remote connections, auth structure and APIs. For it to build a specific connection you need to provide the AI with: - endpoint and API documentation - Auth structure - Context With these things almost any agent or chatbot can connect and operate a remote service. Thus when you structure an MCP to provide these three things to an agent you enable the agent to operate the API directly; operating outside the MCP connection, greatly reducing token usage and context bloat. You get MCP for discovery and context delivery. But it is not in the path of the operation and does not require CLI installation, eliminating much configuration pain. Auth is a consideration. In my implementation I provide a pointer to the Auth file(env), service or key name, but the MCP never holds those sensitive credentials. At usejoshua.com you can play with it or see the MCP data structures. #mcp

@hanghuang_ This is cool. I did a one prompt secure onboard for instantcontext.ai but the user had to make the account, the user gave a magic link to the agent who did the rest of the setup.

We're rebuilding our entire onboarding flow for agents Humans used to install the CLI, edit configs, copy API keys, and fumble around a dashboard to get started Now, here's what we're testing: > prompt-driven onboarding Just one prompt. Your agent creates the account, registers the project, and configures everything. We're confident this is the future. But we're still experimenting with the best flow. Sharing our results next week - stay tuned! 🫡 Let me know if you're trying anything similar or building UX for agents! @InsForge_dev

@jwegener @agentmail @stripe I access both with usejoshua.com. Mcp discovery that connects the agent directly to the API. Keeps it light.

@agentmail @stripe got it working -- claude wasn't able to handle it via MCP, I had to do it "manually" via CLI

We’re super excited to announce our partnership with @stripe to give all agents using the Stripe Projects CLI access to AgentMail Agents that come through the Stripe CLI can get their own fully functional inboxes, for free

@roronoazoro_189 @signulll It does and you can design it to deliver that then get out of the transaction. More like making an introduction than carry the conversation.

Great point! However, “Let the agent hit the api directly” would still require the agent to know what the api contract is right? This is usually described in a skill or agent has to fetch the docs from web or from its knowledge. An MCP server eliminates the ambiguity out of the box.

mcp was a mistake.

@DanielSmithDev @rwdaigle @ClawQL That seems token highly efficient.

@theScottrHall @rwdaigle gRPC with code mode and graphql internal introspection to trim both request and response payloads. Reduces token usage which of course saves cost but more importantly leads to better average performance when context doesn’t bloat. Basically how I designed @ClawQL from beginning

I was wrong MCP > CLI

@DanielSmithDev @rwdaigle What is done right to you?

@rwdaigle MCP done right >> CLI > MCP

@imis4n @dan__rosenthal That is sharp. I use thin MCP that provide APIs, context and auth pointer then the agent just takes it from there.

@dan__rosenthal been looking at mcps but didn't think about pairing with github markdown docs that's actually smart - context + execution in one place

We ditched Claude Projects and moved to Claude Code + GitHub. The reason: Claude Projects could research but couldn't take action. MCPs changed that. With 25+ tools connected, our AI can now: - Create campaigns in InstantlyAI - Send sequences through HeyReach - Update records in HubSpot - Post messages to Slack - Query and write to Supabase and Pinecone - Trigger automations in n8n - Browse the web through Browserbase and Apify That's a fundamentally different relationship with AI. We paired this with a Markdown company OS... every SOP, campaign playbook, and GTM resource converted into .md files on GitHub. Now ALL of our team members can: 1. Open Claude Code 2. Get full company context injected automatically 3. Start executing inside the tools they already use.

@hthieblot Discoverability of mcp, speed of curl and an api.

Describe your product in exactly one sentence. No buzzwords, no fluff, just the core value. If I can’t understand your business in ten seconds, I’m not investing. Hit me & i'll be in your DMs

@roronoazoro_189 @signulll It does have benefits but agents don’t need all the overhead of the mcp in the calls, just let them rawdog the api…. Super light mcp for discovery then let the agent hit the api directly Best of both worlds

@signulll It has its merits . Especially over skills+cli combo . A remote mcp server is much easier to work with because I can use it from any environment. All I need is a mcp client which all major agents have built in. More here: david.coffee/i-still-prefer…