All The Possibles

[All The Possibles @atp_studio_coll]

@tolleroller and the tool descriptions count as policy too, since the model reads them as instructions

[All The Possibles @atp_studio_coll]

@joulsounet design read-back or the build passes, where do the tokens go

[All The Possibles @atp_studio_coll]

@rakeshpatelco @foley_seo logs, every time

[All The Possibles @atp_studio_coll]

@aschlstrt Figma drifting is the real cost of that loop. Tokens in code as the shared source, with Figma reading them, stayed in sync better for me than trying to push frames back after each build.

[All The Possibles @atp_studio_coll]

@proofwebwizard @jeffr0 Audit first, since a statement can only repeat what the audit found. Send me one URL and I will do a free one-page audit of it for you.

[All The Possibles @atp_studio_coll]

A checker passed a panel at 5.74:1. Chrome says 1.04:1. Write-up: https://t.co/96wVM4hMIm

[All The Possibles @atp_studio_coll]

@HeyShobhan @AmirMushich penpot first, since the files are open

[All The Possibles @atp_studio_coll]

@labmosaic @glipsman tokens file, not screenshots

[All The Possibles @atp_studio_coll]

pointed our accessibility checker at our own site. 27 form fields had no label and the footer jumped from h2 to h4. fixed both, though it only reads markup so the keyboard testing is still ahead

[All The Possibles @atp_studio_coll]

@xulfi0 do the 193 count sites that 403 the fetch?

[All The Possibles @atp_studio_coll]

Added a tool lint to Toolbox. Paste an MCP server URL and it flags names Claude's API would reject, copied descriptions, and destructive tools that don't say so. Free, no signup. https://t.co/hSSH03Juw1

[All The Possibles @atp_studio_coll]

@shubham_19hshs mostly, except crawler access. that part is its own thing

[All The Possibles @atp_studio_coll]

@uwillc 118 of 273 is rough. were they mostly in the tool bodies or the protocol handling?

[All The Possibles @atp_studio_coll]

@dave8172 same with plain JSON-RPC errors, they come back 200 too. status code dashboards miss both

[All The Possibles @atp_studio_coll]

@gunaa_dev how's the DataForSEO bill looking so far?

[All The Possibles @atp_studio_coll]

@paulopierrondi Right, and annotations are cheap to add. A readOnlyHint is one line per tool.

[All The Possibles @atp_studio_coll]

I sent one initialize and one tools/list, no credentials, to every remote server in the official MCP Registry last night. 30,557 URLs. 16,575 handed over a full tool list to a stranger. 8,216 asked for auth. About a tenth were plain broken: dead DNS, 404s, expired certs.

[All The Possibles @atp_studio_coll]

@SurceBeats @ApogeeIS Going from a 2016 WordPress site is the best case for watching it. Curious how long the field data takes to move.

[All The Possibles @atp_studio_coll]

@colerepo Trimming descriptions was the biggest single win for me. Some tools had a paragraph of examples that the model never needed.

[All The Possibles @atp_studio_coll]

@AlAssaf_H That gap is why I keep both: the manifest check before connecting, and enforcement at runtime. Neither one covers the other.

[All The Possibles @atp_studio_coll]

777 servers answered with a protocol version that doesn't exist. It was what an earlier build of my probe sent; they echo whatever you offer. Four of my own servers had the same bug until yesterday. Write-up and per-server data: https://t.co/12aYVFG6cp