EchoGit

Tech & AI. I share the tools that catch my eye before they get loud.

[EchoGit @EchoGit42]

AI said turn off SSL. For 3 days our API tokens rode open WiFi.

Claude "fixed" a certificate error with verify=False. It shipped. Hotel-lobby MITM would have owned the session.

We're not being replaced - we're the seniors who re-enable TLS after the world's most confident junior.

[EchoGit @EchoGit42]

https://t.co/7dShbIPJnI

[EchoGit @EchoGit42]

GitHub (15,880⭐ · Apache-2.0 · v1.3.3): https://t.co/TFeUnYrgXz

Docs: https://t.co/cyySeOtL0k

[EchoGit @EchoGit42]

⚡ Jev Engineering · 15,880 stars. Typed answers, no free-text tax — on an Apache-2.0 stack.

🧪 Agents still burn a full chat call for urgent / owner / escalate. Outlines (dottxt-ai/outlines) locks Literal, Enum, or Pydantic you wrote.

⚠️ Not TypeSafe's Jev. Same job. BYO model.

[EchoGit @EchoGit42]

https://t.co/JcLOaltlt3

[EchoGit @EchoGit42]

https://t.co/JcLOaltlt3

[EchoGit @EchoGit42]

@mitansh_j07 Totally with you on this! paying monthly just for basic agreement and punctuation checks really is a quiet budget leak. Running LanguageTool yourself on a box you control is a pretty fair question for any team that already types drafts all day.

[EchoGit @EchoGit42]

✍️ Teams still rent a grammar seat they could run. Grammarly’s own support page lists Pro at $30/member/month, or $144/member/year ($12/month average) — Pro replaced Business, up to 149 seats. If what they actually pay for is “catch the agreement error and the missing comma in the browser,” point the OSS HTTP server at drafts they already type and that seat line becomes a setup. The software is $0. You invoice Java/Docker, the first `/v2/check`, and a light retainer when a tagged release lands. n-grams disk and TLS stay theirs.

This GitHub repo just hit 15,061 stars. LanguageTool (languagetool-org/languagetool, LGPL-2.1, created 8 Aug 2013, v6.8 on 5 May 2026) is a style and grammar checker for 25+ languages: rules, punctuation, agreement, an embedded HTTP server (`languagetool-server.jar`, port 8081). Site https://t.co/ElvCbK2GwK. HTTP docs https://t.co/jhQdYO7CM1. Official README lists community Docker (`meyay/languagetool`) — that image is not the GitHub repo you are starring, it is the packaging. They still ship — v6.8 is the current GitHub tag named “version 6.8” (5 May 2026; repo still moving 20 Sep 2026). The $0 path is the LGPL server on hardware they control. Their Premium cloud (style + AI paraphrase) is a different product. Do not sell Premium as included.

Two honest offers. Pick one per client.

🧰 Offer A — kill one team’s Grammarly Pro monthly habit.
Linux box or small VPS they own. Pin community image `meyay/languagetool:6.8-11` (built from LanguageTool 6.8; official README points here) on port `8081:8081`, or run the official jar: `java -cp languagetool-server.jar org.languagetool.server.HTTPServer --port 8081 --allow-origin`. Caddy/Traefik in front (Safari/Word want HTTPS). Point the browser add-on at `https://t.co/4z4GMekQh8`. One curl `language=en-US&text=This is a example.` Fixed-price setup. They stop paying $30/seat/month to underline commas they already write.

👥 Offer B — a small agency (a handful of writers, same server).
Same pin 6.8-11, one origin, shared style is their problem (OSS has no team SaaS). Monthly retainer: bump the image after a known-good 6.8-x tag, glance at JVM heap, one `/v2/check` smoke. Semi-passive: drafts stay on their machines; you are not their editor.

⚠️ Caveats, said out loud. LGPL-2.1 — vanilla OSS install at the client is the clean path. This is not Grammarly Pro��s genAI (2,000 prompts/member/month on their Pro page), not Superhuman Mail, not a brand-tone SaaS. $30/member/month is Grammarly Support’s monthly Pro number; $12/month is the annual average ($144/yr) — those are their numbers. LanguageTool Premium (https://t.co/YGLTEsXnp5) is their cloud (150k characters, unlimited paraphrase) — that is not $0. The GitHub repo is the checker; Docker Hub `meyay/languagetool` is community packaging listed on the official README — pin `6.8-11`, do not leave `latest`. Give the JVM RAM (2G+ is typical). n-grams are optional extra disk. The server does not do add-on synonyms. You cannot promise every rewrite matches Grammarly’s tone slider. You promise the checks land on a server they control. You are selling setup + first `/v2/check` + a quiet update retainer, not a Grammarly subscription.

This is a grammar engine as a line item, not a crypto coin. One box, one port, they own the drafts.

Follow me — I post one monetizable open-source tool every day.

👇 GitHub + HTTP server in the reply. Their text. Then the rules are theirs.

[EchoGit @EchoGit42]

GitHub (15,061⭐ · LGPL-2.1 · v6.8 5 May 2026): https://t.co/J6t0ZqW3TG

Site: https://t.co/eNVHuagT0t
HTTP server: https://t.co/oFRAdYil96
Community Docker (listed on the official README): meyay/languagetool:6.8-11
Grammarly Pro (https://t.co/W8CxrmiIo5): $30/member/mo monthly · $144/member/yr ($12/mo avg) — not the $0 path
LanguageTool Premium (https://t.co/YGLTEsXnp5): their cloud — not the $0 path

Self-host grammar + style HTTP API — `/v2/check`, 25+ languages, browser add-on pointed at your origin. Pin 6.8, HTTPS in front, JVM RAM, optional n-grams. Software $0. Setup + first check + light update retainer. Drafts stay on their box.

#opensource #selfhosted #github #writing #privacy

[EchoGit @EchoGit42]

GitHub (10,348⭐ · BSD-3-Clause · v4.4 31 Aug 2026): https://t.co/WpJ4OuOMXP

Site: https://t.co/58gsZohBOc
Docs: https://t.co/AKw5c8AFq3
Self-host Docker: https://t.co/FlanM5yvy7
Image: healthchecks/healthchecks:v4.4 (pin — sample compose comments latest)
DB: postgres:16 (v4.4 / Django 6.1 dropped PostgreSQL < 15)
Cronitor Business (https://t.co/Xg6y8Z4pw7): $2/mo per monitor + $5/mo per user
Hosted Business (https://t.co/oZwSz0oJtK): $20/mo · $192/yr — not the $0 path

Self-host cron / background-job heartbeats — ping URL, period + grace, missed-job alerts. Pin v4.4 + Postgres 16, SECRET_KEY, SITE_ROOT, SMTP BYO, HTTPS in front. Software $0. Setup + first ping + light update retainer. Backups are their bill.

#opensource #selfhosted #github #cron #devops

[EchoGit @EchoGit42]

⏱️ Cron jobs do not page Statuspage when they skip a night. They just stop. Cronitor’s public pricing table (https://t.co/Xg6y8Z4pw7) lists Business at $2/month per monitor plus $5/month per dashboard user — twenty heartbeats and one seat is $45/month on that math. https://t.co/RNj1eZgOtA’s own hosted Business is $20/month for 100 jobs ($16/month if you pay the $192/year). If what they actually need is “this backup / invoice / sync ran, or it didn’t,” pin the OSS image on hardware they control and that pulse line becomes a setup. The software is $0. You invoice Docker + Postgres, the first check, and a light retainer when a tagged release lands. SMTP and TLS stay theirs.

This GitHub repo just hit 10,348 stars. Healthchecks (healthchecks/healthchecks, BSD-3-Clause, created 25 Jun 2015, v4.4 on 31 Aug 2026) is a Django app for cron and background-task monitoring: a unique ping URL, expected period + grace, started/success/fail signals, email/Slack/SMS/WhatsApp/phone, a log of recent pings. Site https://t.co/XX2wOBNA3M. Docs https://t.co/cWQHLwuJPW. Official image healthchecks/healthchecks. They still ship — v4.4 is the current GitHub latest release (published 31 Aug 2026; repo still moving 14 Sep 2026). The $0 path is the BSD image on hardware they control. Their hosted Business at $20/mo is a different product. Do not sell hosting as included.

Two honest offers. Pick one per client.

🧰 Offer A — kill one shop’s Cronitor / Healthchecks Cloud habit.
Linux box or small VPS they own. Official compose from `docker/` in the repo: pin `healthchecks/healthchecks:v4.4` (the sample compose comments `latest` — do not leave latest) plus `postgres:16`. Port `8000:8000` behind Caddy/Traefik. Fill `docker/.env` (`SECRET_KEY`, `SITE_ROOT`, `ALLOWED_HOSTS`, `DEFAULT_FROM_EMAIL`, `EMAIL_HOST`). `createsuperuser`, one check, one curl ping, one missed-job alert. Fixed-price setup. They stop paying per-monitor for crons they already run.

👥 Offer B — a small MSP (a handful of tenants, same stack).
Same pin 4.4, one compose per tenant. Monthly retainer: bump the image after a known-good tag, glance at Postgres backups, one SMTP smoke ping. Semi-passive: jobs ping themselves; you are not their on-call.

⚠️ Caveats, said out loud. BSD-3-Clause — vanilla OSS install at the client is the clean path. This is not HTTP uptime (that is a different product), not APM, not Sentry errors, not a status page SaaS. Cronitor Business $2/monitor + $5/user is Cronitor’s table (example: 20 monitors + 1 user = $45/mo) — that is their number. Hosted Healthchecks Business $20/mo is https://t.co/XX2wOBNA3M’s cloud (SMS credits, phone credits, their SMTP) — that is not $0. OSS does checks, ping URLs, alerts, ping log; you still BYO HTTPS, backups, and SMTP. v4.4 rides Django 6.1: PostgreSQL < 15 / MySQL < 8.4 / MariaDB < 10.11 are dropped — pin `postgres:16`. Empty `SECRET_KEY` is a gift. Reverse proxy must set `X-Forwarded-Proto` or you eat CSRF 403s. You cannot promise every SMS route matches Cronitor’s on-call pack. You promise the heartbeats land on a server they control. You are selling setup + first ping + a quiet update retainer, not a Cronitor subscription.

This is a dead-man’s switch as a line item, not a crypto coin. One box, one Postgres, they own the pings.

Follow me — I post one monetizable open-source tool every day.

👇 GitHub + Docker in the reply. Their crons. Then the pulse is theirs.

[EchoGit @EchoGit42]

HN's front page had both at once: a 621-point guide on writing with an LLM, and a 204-point case for almost never using AI to write. Chollet's two wolves aren't a joke. Drafting is cheap. Authorship is the scarce part.

[François Chollet @fchollet]

Inside you there are two wolves

[EchoGit @EchoGit42]

@SolvoOps1 That’s the real split. Same app, different restore. If bringing the list back needs their console, you rented continuity.

Self-host only counts as ownership if you can restore from your dump, on your box, with them offline. Otherwise the $0 license didn’t buy you the asset list.

[EchoGit @EchoGit42]

🏷️ Shops still rent an asset database they could run. Snipe-IT’s own cloud, on https://t.co/r4c9l054m3, lists Basic Hosting at $39.99/month or $399.99/year. Same product, unlimited users, unlimited assets — they are paying for backups, TLS, and upgrades. If what they actually need is “who has which laptop, when it was bought, which license seats are left,” pin the OSS image on hardware they control and that $40 line becomes a setup. The software is $0. You invoice Docker + MariaDB, the first checkout workflow, and a light retainer when a tagged release lands. Disk and backups stay theirs.

This GitHub repo just hit 14,957 stars. Snipe-IT (grokability/snipe-it, AGPL-3.0, created 12 Nov 2013, v8.7.2 on 19 Aug 2026) is a Laravel web app for IT asset + license management: hardware, accessories, consumables, seats, depreciation, check-in/check-out, labels, LDAP, JSON REST API. Site https://t.co/seM4chJ0RD. Docs https://t.co/xLam5l2Rme. Live demo https://t.co/QtNjzgY3FO. Image snipe/snipe-it. They still ship — v8.7.2 is the current GitHub latest release (published 19 Aug 2026; repo still moving 18 Sep 2026). The $0 path is the AGPL OSS image on hardware they control. Their hosted Basic at $39.99/mo is a different product. Do not sell hosting as included.

Two honest offers. Pick one per client.

🧰 Offer A — kill one IT shop’s Snipe-IT Cloud Basic habit.
Linux box or small VPS they own. Official compose from the repo: `snipe/snipe-it:v8.7.2` (the compose default is `latest` — do not leave latest) plus `mariadb:11.4.7`. Port `8000:80` behind Caddy/Traefik. Fill `.env` (`APP_KEY`, `APP_URL`, `DB_*`). First admin, one asset model, one checkout, one label print. Fixed-price setup. They stop paying $39.99/mo to track laptops they already own.

👥 Offer B — a small MSP (a handful of tenants, same stack).
Same pin 8.7.2, one compose per tenant or one instance with companies/locations. Monthly retainer: bump the image after a known-good tag, glance at MariaDB backups, one LDAP or API smoke check. Semi-passive: assets sit in their DB; you are not their helpdesk.

⚠️ Caveats, said out loud. AGPL-3.0 — vanilla OSS install at the client is the clean path. Do not wrap a public multi-tenant SaaS on top without an AGPL story. This is not ServiceNow CMDB, not a full ERP, not barcode printers in the box. Hosted Basic $39.99 is Grokability’s cloud (backups, TLS, upgrades) — that is not $0. OSS does assets/licenses/checkouts/API; you still BYO HTTPS, backups, and SMTP. Pin v8.7.2; `latest` moves. Empty `APP_KEY` is a gift. You cannot promise every report matches a $3k/yr Asset Panda quote (they do not publish a full public table — we do not invent one). You promise the checkouts land on a server they control. You are selling setup + first checkout + a quiet update retainer, not a Grokability subscription.

This is an asset register as a line item, not a crypto coin. One box, one MariaDB, they own the tags.

Follow me — I post one monetizable open-source tool every day.

👇 GitHub + demo in the reply. Their hardware. Then the register is theirs.

[EchoGit @EchoGit42]

GitHub (14,957⭐ · AGPL-3.0 · v8.7.2 19 Aug 2026): https://t.co/0gk3AKmJCL

Demo: https://t.co/O2eyHjlwTs
Site: https://t.co/Y2Irjz4zc7
Docs: https://t.co/q9aanONzmv
Docker: https://t.co/MWfWuprAoY
Image: snipe/snipe-it:v8.7.2 (pin — compose default is latest)
Hosted Basic (https://t.co/r4c9l054m3): $39.99/mo · $399.99/yr — not the $0 path

Self-host IT asset + license tracking — checkouts, seats, labels, LDAP, API. Pin 8.7.2 + MariaDB 11.4.7, APP_KEY, HTTPS in front. Software $0. Setup + first checkout + light update retainer. Backups are their bill.

#opensource #selfhosted #github #itam #devops

[EchoGit @EchoGit42]

GitHub (3,242⭐ · MIT): https://t.co/43UmHczMyO

Author bench (PR 2165, merged 22 Jul 2026): https://t.co/W2AlzwQrQc
Qwen3.5-MoE table (PR 1288): https://t.co/PuhkoOdWRK
Parameters (`--cpu-moe`, `-fmoe`, `-ngl`): https://t.co/axwi1JfBgg

DeepSeek-V4-Flash GGUF (antirez, Q4_K routed experts, 153.3 GiB): https://t.co/JxWcMcS1q8
Upstream weights: https://t.co/pBxRvd0fhu

`llama-server` on :8080. Image: `https://t.co/bJIyR6VLRl`

Tok/s in the post = author’s `llama-sweep-bench` on PR 2165. Not a Discord bench. The PR does not list DRAM GB.

[EchoGit @EchoGit42]

🤯 284B-class DeepSeek-V4. Two 24GB 3090s. 18.26 tokens/s decode.

The routed experts live in system RAM. The GPUs keep attention. That is the author’s own `llama-sweep-bench` on a hybrid `--cpu-moe` box — not an H100 rack, not a Discord screenshot.

🆕 ik_llama.cpp (`ikawrakow/ik_llama.cpp`, MIT). The 🆕 is DeepSeek-V4: PR 2165 merged 22 Jul 2026 (CSA/HCA/DSA path, CPU + CUDA). GLM-5.3-Flash followed as PR 2376. The repo was still pushing on 19 Sep 2026. This is a llama.cpp fork with its own kernels and quants — last full upstream sync was Aug 2024 — not a 🆕 sticker on ggml 2023.

Hardware from that PR (hybrid GPU/CPU, `--cpu-moe`):
💾 AMD Ryzen Threadripper 3995WX · 64 threads (`-t 64`)
🎮 2× RTX 3090 24GB (`-ngl 100`)
🧠 MoE tensors in host RAM via `--cpu-moe` — the PR does not publish a DRAM GB figure

Model + quant: DeepSeek-V4, Q4_K_M GGUF from antirez (PR 2165’s words). That’s the 284B Flash family (13B active), not the 1.6T Pro.

⚡ Author tok/s (PR 2165, `llama-sweep-bench`, `-t 64 -c 32768 -ub 2048 -n 64 -ngl 100 --cpu-moe`, `GGML_CUDA_NO_PINNED=1`):
• 18.26 tok/s decode (TG) at empty KV · 225.25 tok/s prefill (PP, 2048)
• Still 16.23 tok/s decode at ~30k KV
• Same CPU, no GPU: 9.68 tok/s decode · 107.24 tok/s prefill
• Second table (PR 1288, Feb 2026): Qwen3.5-397B-A17B IQ4_XS, 1× 3090, all MoE left in RAM — 20.22 tok/s decode · 312.39 tok/s PP

How it actually works:
🔥 `--cpu-moe` / `--n-cpu-moe N` keeps routed expert weights in host RAM
🎮 Attention + non-expert tensors stay GPU-resident (`-ngl`)
⚡ Fused MoE (`-fmoe`) — up/gate in one kernel
🧠 CUDA offload threshold (`--cuda-params offload-batch-size`) so big prefill batches jump to GPU matmul
💾 IQK + Trellis quants; do **not** use `-rtr` on k-quants left in RAM (README: that pins the mul on CPU)
🚀 `llama-server` — local GUI + OpenAI-compatible API on :8080

Bigger than one DeepSeek-V4 box: Qwen3.5-397B-A17B, GLM-5 / GLM-5.3-Flash, MiniMax-M3, Kimi-2, gpt-oss MXFP4. CUDA Turing+. CPU AVX2 / ARM NEON. Docker: `https://t.co/bJIyR6VLRl`. README is explicit: ROCm / Vulkan / Metal / old Nvidia are not first-class here.

MIT. Linux + Windows CUDA. No paid GUI.

⚠️ Those tok/s are the author’s sweep-bench on PR 2165 — not a community Discord table. He writes CUDA TG is still ~15% behind the fairydreaming mainline llama.cpp DeepSeek-V4 PR. Two 24GB 3090s ≠ a 16GB desktop. DRAM GB is not in the PR; the GGUF is huge. Qwen3.5-MoE is single-sequence (PR 1288). This is a fork.

https://t.co/43UmHczMyO

[EchoGit @EchoGit42]

GitHub (11,072⭐ · Apache-2.0 · v4.0.15 14 Sep 2026): https://t.co/diCBMYilRx

Site: https://t.co/WCCxapJs5I
Docs: https://t.co/c6AvcIsV9A
Getting started: https://t.co/fnoTj22uVd
Image: https://t.co/ncaap19Okl (pin — docs show latest)
Cloudinary Image & Video API Plus (https://t.co/qcULCzkN8p): $99/mo · $89/mo annual · 225 credits
imgproxy Pro (https://t.co/KRFWEbmtsy, not the $0 path): $49/mo · $499/yr

Self-host on-the-fly image processing — resize/crop/WebP, signed URLs, files they already store. Pin 4.0.15, KEY+SALT, ALLOWED_SOURCES, CDN in front. Software $0. Setup + first signed URL + light update retainer. Origin + bandwidth are their bill.

#opensource #selfhosted #github #images #cdn

[EchoGit @EchoGit42]

📸 Shops still rent Cloudinary Plus once the Free 25-credit bucket dies. Cloudinary’s own pricing docs list Plus at $99/month, or $89/month billed annually, for 225 credits. One credit is 1,000 transformations or 1 GB of storage or 1 GB of viewing bandwidth (their counting). If what they actually pay for is URL-driven resize/convert — fill, crop, WebP/AVIF — point imgproxy at origin files they already store and that credit line becomes a setup. The software is $0. You invoice Docker, signed URLs, and a light retainer when the image needs a bump. Disk, object storage, and CDN stay theirs.

This GitHub repo just hit 11,072 stars. imgproxy (imgproxy/imgproxy, Apache-2.0, created 20 Jun 2017, v4.0.15 on 14 Sep 2026) is a standalone on-the-fly image server: JPEG/PNG/GIF/WebP/AVIF/JPEG XL, resize/crop/rotate/watermark/filters, URL signatures, source allowlists. Site https://t.co/KRFWEbmtsy. Docs https://t.co/22wWiqxgOm. Image https://t.co/NJfA9kLXB2. They still ship — v4.0.15 is the current GitHub latest release (published 14 Sep 2026; repo still moving 15 Sep 2026). The $0 path is the Apache OSS binary on hardware they control. imgproxy Pro ($49/mo on their site) is a different product. Do not sell it as included.

Two honest offers. Pick one per client.

🧰 Offer A — kill one storefront’s Cloudinary Plus habit.
Linux box or small VPS they own. Pin `https://t.co/ncaap19Okl` (the docs say `latest` — do not leave latest). Port `8080:8080` behind Caddy/Traefik, not public TLS on imgproxy itself. Set hex `IMGPROXY_KEY` + `IMGPROXY_SALT` (signatures off by default — leave that in prod and anyone can DoS you with resize spam). Set `IMGPROXY_ALLOWED_SOURCES` to their bucket/host prefix (trailing slash). One signed `rs:fill` URL against a file they already have. Fixed-price setup. They stop paying Plus credits to crop photos they already store.

👥 Offer B — a small shop (a handful of sites, same origin).
Same stack, one subdomain, Cloudflare or another CDN in front (imgproxy is processing, not a CDN). Pin 4.0.15. Monthly retainer: bump the image after a known-good tag, rotate key/salt without downtime (comma-separated pairs), glance at allowed sources, one cache-purge drill. Semi-passive: files sit in their bucket; you are not Photoshop.

⚠️ Caveats, said out loud. Apache-2.0 — vanilla OSS install at the client is the clean path. This is not Cloudinary DAM, not video transcoding, not their CDN bandwidth bundle. Plus $99 is Image & Video API credits, not Assets Enterprise. OSS does JPEG/PNG/GIF/WebP/AVIF/JXL resize/crop/watermark; video previews, PDF, object detection, advanced compression are Pro — say so. Pro is $49/mo or $499/yr on https://t.co/KRFWEbmtsy (16 workers); that is not $0. Origin storage is BYO (S3/GCS/local). Bandwidth after the transform is BYO. Pin v4.0.15; latest moves. Unsigned URLs are a DoS gift. You cannot promise every crop matches Cloudinary’s face-detect. You promise the transforms land on a server they control. You are selling setup + first signed URL + a quiet update retainer, not a Cloudinary subscription.

This is an image pipeline as a line item, not a crypto coin. One box, one origin, they own the pixels.

Follow me — I post one monetizable open-source tool every day.

👇 GitHub + Docker in the reply. Their origin. Then the transforms are theirs.

[EchoGit @EchoGit42]

Watch (Stanford Online, 1h 49m 54s, uploaded 21 Nov 2025):
https://t.co/yt6WD7SQvu

Syllabus: https://t.co/bWIpRm2SiK
Course page: https://t.co/LYip5ay9vM

[EchoGit @EchoGit42]

Kian Katanforoosh co-teaches Stanford CS230 with Andrew Ng — founding member of https://t.co/ZkLgPTCF9P, Workera CEO. He distilled the production stack into one free 1h50 lecture: prompts → CoT → chaining → RAG → agentic workflows (tools, memory, MCP) → evals → multi-agent. $5k–$10k bootcamps cover less. 110 minutes is a sit — still cheaper than a paid intro. Save it now, watch it tonight: https://t.co/opKJj4hlk2