code mode is the one to watch. does the output of each call inside the script stay out of the context, or does claude still read every result? in our experience at ApyHub the tool list is the other cost: the whole catalog sits behind 4 tools, so the agent only reads the spec it picks
@shabirkhan_dev@apyhub: an API catalog your agents can call over MCP, just by asking Claude, no glue code. saw you're building Grid, how are you handling shared context across agents?
Notion’s AI chat transcripts may have been exposed. The cause was a breach at Fireworks AI, a model provider it used.
Every vendor in your stack is part of your attack surface. 🤔
How do you keep track of them on small projects?
#APIs#DevelopersIndia
Every API you publish on @apyhub is MCP-ready by default.Claude, Cursor and other agent runtimes can discover and call it. You write no wrapper, no tool definition and no glue code.Bring an endpoint. We handle the rest.#MCP#AIAgents#API
🚀 Decode and Validate VIN API on @apyhub!
VIN Decoder validates a vehicle identification number and returns decoded vehicle details from the VIN.
Try it ↓
#API#DeveloperTools https://t.co/nPWFq1p9Al
@antigravity knowledge work is where agents burn time on file chores: pdf to text, docx to pdf, tables out of scans. in our experience at ApyHub, handing those to MCP endpoints instead of letting the agent write a converter keeps runs short and the output the same every time
When did you last open your ~/.claude or ~/.codex folder? 👀
A GitHub issue says Postman’s desktop app added plugins and a startup hook there without asking. Still open.
Auditing mine tonight.
How do you keep your agent config clean?
#ClaudeCode#DevTools
@_philschmid where does discovery go in code mode? the sandbox can call anything, but the model still has to learn which API exists and what it expects. in our experience at @apyhub that's the step worth keeping as a tool: search first, then read only the spec it picked
apple is putting alibaba's qwen into apple intelligence in china, dev testing starts nov 4 🍎
same feature, different model underneath depending on the country
if your AI provider vanished tomorrow, how long to swap it?
a) an hour
b) a week
c) please don't ask
#AI#iOSDev
@vectorman does it check the final URL after redirects or only the one you send? from our experience at @apyhub agents need a link check before opening anything, and we have no phishing scanner in the catalog yet. could list it, no exclusivity, ~10 min: https://t.co/rOakgvzcFm
@Saboo_Shubham_ point 4 is where the tool list hides too. in our experience at ApyHub, putting the whole catalog behind 4 generic tools (catalog, search, spec, call) means the agent only reads the spec for the endpoint it actually picked
Most API clients open a new request as a form: URL bar, then tabs for Params, Headers, Body, Auth, Scripts, Tests. Some stay empty.
In Voiden you paste the curl from the docs into a .void file and get only the blocks it needs: Endpoint, Headers, JSON body etc.
Need auth? Add that as a new block.
try Voiden here: https://t.co/5s9ObUFrzC
#DevTools #webdev
@rileybrown not the two-host format, but close: our text to speech endpoint turns text or a doc url into an mp3, callable from any agent via the @apyhub MCP. in our experience at ApyHub the agent can write the dialogue itself, the voice is what it needs an API for
@_philschmid --dry-run is the flag we'd reach for first. does it print the exact request, body and uploaded files included? in our experience at @apyhub agents pick the right call more often when they can see it before spending anything
@XRuturaj_ the provider sees whatever is in the context, and it can land in logs and traces too. in our experience at @apyhub the fix is keeping the key out of the prompt: it sits in the MCP config, the agent calls the tool and never reads it
Voiden 2.3.3 is out 📷
→ Reworked Commit History: hover a commit for author, message and files
→ Requests inherit .env files from parent folders
→ Optional request fields
→ Choose what opens on startup
→ Fixes across Git, envs, gRPC and saving
https://t.co/Bv4GwLlF1p
10 MCP servers or 1? 😅
claude code just fixed errors in sessions with lots of MCP tools showing up as "prompt is too long"..
i lean small APIs behind one server (it's what we do at @apyhub)
how many tools does your agent carry?
#MCP#APIs
CLI, MCP or a plain REST API? 🤔 when you give your agent a new tool, which one goes in first?
i'm biased (we serve APIs over REST and MCP at @apyhub) but CLIs seem to be having a moment on github 👀
what do you reach for first, and what made you switch? 👇
#MCP#AIagents
@alliekmiller and when the tool does have an API, hand the agent that first. in our experience at @apyhub a call comes back in the same shape every run, a recorded click path breaks the day someone moves a button
For debugging by hand, @VoidenMD lets you change one block (headers, auth, params) at a time and resend, and keeps the working request in your repo.
The full guide, with the browser fixes and the nginx config:
https://t.co/bFKpNloQqx
Retrying a 400 Bad Request never works.
The same request fails the same way, every time. Something has to change first.
Here are the 6 things behind almost every 400, and a 5-step check that finds yours 🧵
AI agents hit 400s constantly. They guess field names, send numbers as strings, then retry the same broken request and burn tokens.
ApyHub MCP gives Claude, ChatGPT, Cursor and Le Chat every API as a tool with a fixed schema, so there's nothing to guess.