Ben Thompson says agents make everyone a hacker, and Apple's garden a prison.

Strata (MIT, C++, 12,523 stars, 845 commits, 68 contributors, created September 24, v0.1.39 on Sunday) runs Qwen3.8-Flash-Next, a 125B mixture-of-experts model, on a Windows or Linux gaming PC with 12 GB of VRAM and 32 GB of RAM. The model has 24,576 experts and uses 10 per token, so Strata keeps the busiest experts on the graphics card, all of them in RAM, has the processor work on the rest at the same time, and leaves a 29 GB lookup table on the SSD. A small draft model guesses ahead and the big one checks, which the README credits with 1.6-1.8x. On the README's RTX 5070 with 64 GB of RAM, it writes 94 tokens per second at Q2_0 and 53 at IQ3_S, and reads a 32K prompt at 1,620 to 2,650 tokens per second. It serves OpenAI's /v1, the Responses API and Anthropic's /v1/messages, so ANTHROPIC_BASE_URL=http://127.0.0.1:8080 points Claude Code at it. Windows users double-click START-HERE.bat, which downloads about 70 GB.

The HN thread (826 points, 365 comments) was mostly about what those speeds cost. The sizes are rated "good", "better", "great" and "best"; only IQ3_S is said to match the full model on published tests, and the Coder build drops half the experts to reach 91% of the full model's SWE-bench Verified score, measured by the people who pruned it. One commenter cited a paper finding that 2-bit quantization "often causes broad degradation", another said these projects "conveniently" leave accuracy out of their speed numbers, and a third, running Q1 and IQ2 quants in llama.cpp, saw the model invent URLs. On the other side, a user switched to IQ3_XXS and reported 60 tokens per second on Strata against 21 in llama.cpp, with better output. Asked why the expert caching isn't in llama.cpp, a commenter pointed to its rule that a human must understand any contributed code, which keeps AI-written kernels like these in their own forks.

Reach for it if you have a 12-24 GB card, 64 GB of RAM, and coding work you'd rather keep off a vendor's servers. Delete the cloud key your agent uses for the tasks you'd run locally if a local model were good enough. Tradeoffs: the fast sizes are 2- and 3-bit quants with no accuracy numbers beside the speed table; the PC can stop responding for 1-3 minutes while 35-55 GB loads; it answers one request at a time by default; and the suggested install is to paste the repo URL into your coding agent and let it follow docs/AI_SETUP.md.

RemoveMacAI (MIT, Swift, 1,054 stars, 12 commits, one author, created September 29, v0.2.4 on Monday) exists because macOS 27 has no single switch for Apple Intelligence, and its models stay on disk after you turn the features off. It installs a configuration profile with Apple's restriction keys, forces the settings that have no key, removes the models through Apple's asset service, and points each model's download at a closed local port so macOS doesn't fetch it again. System Integrity Protection stays on and nothing under /System is touched. removemacai status shows each feature and the model sizes, off --keep <features> leaves some on, --dry-run shows the changes, and revert removes the profile. The README credits pared, an earlier tool by 4evy, with first mapping the asset service and model sets.

In the HN thread (623 points, 417 comments), commenters put the reclaimed space anywhere from 12 GB to 30 GiB, and one called it "sitting on my disk against my consent". The top comment objected to the curl | bash install. The README answers part of that: the script checks a SHA-256, runs from a temporary directory, and each release carries a GitHub build-provenance attestation you can verify with gh attestation verify. A sharper objection was why this is an app at all, when it mostly generates a profile: a downloadable, pre-generated .mobileconfig you could read before installing would ask for less trust.

Reach for it if you're on a 256 GB Mac running macOS 27 and don't use the features. Delete the list of defaults write commands you were going to paste one at a time. Tradeoffs: it also disables apps built on Apple's Foundation Models framework, the Shortcuts "Use Model" action, Visual Intelligence and natural-language editing in Calendar; Storage settings keeps counting the models until macOS gets round to deleting the files; 27.0.1 needs 0.2.3 or later, so expect point releases to break it; and it is one author's twelve commits.

SCM (MIT, JavaScript, 327 stars, 9 commits, one author, created Saturday) is an Electron app for Apple Silicon that indexes folders of photos and video and searches them five ways: by meaning (CLIP or SigLIP), by scene inside a video with a jump to the timecode, by visible text (Tesseract), by spoken words (Whisper), and, if you turn it on, by asking a local chat model (Qwen3 1.7B by default) about what the other four extracted. Everything runs on the Mac; weights download once. Watched folders import automatically, files are deduplicated by SHA-256 so renames don't create copies, and switching vision models re-embeds the library in the background while keyword search keeps working. Video density is a preset, from one sample every 60 seconds (Eco) to every 2.5 (Ultra Pro), and the app shows the time and disk cost before you commit.

The Show HN thread (154 points, 69 comments) asked the practical question first: how long 12,000 videos take. SCM doesn't publish a number. A commenter who built something similar with CLIP on an M1 said one frame a second over 12,000 videos took days, and keyframes only got it down to overnight. The default CLIP model takes about half a second per image on the CPU, by the README's own table; SigLIP-2-B/16 takes 50-100 ms. Another commenter said Apple's Vision framework "smokes tesseract" on a Mac, and the author agreed, saying Tesseract was chosen to keep the inference layer portable JavaScript.

Reach for it if you have a footage or screenshot archive outside Apple Photos, on an external or network drive, and you want to find a shot, not a file. Delete the folder-naming scheme you invented so you could find a clip again. Tradeoffs: on a large video library, plan for an overnight first index; the Homebrew cask clears the macOS quarantine flag on every install and upgrade, which means skipping Gatekeeper for an app that is nine commits old; and it is Electron, which several commenters held against it.

Routine Cleanup (MIT, Python, 2 stars, 5 commits, one author, v1.1.0 on Saturday) is a Claude Code skill for a problem anyone with an hourly routine in the Claude desktop app has: every run leaves a session, which comes to 700 or more a month, and the app deletes them one click at a time. The author had 618 in one routine. A planner script, standard library only, sorts each run into autonomous (nobody typed after the kickoff), engaged (you typed something) or unverifiable (the transcript is gone), and by default selects only the autonomous ones, never the newest N, the current conversation, or anything active in the last 15 minutes. The planner can't delete anything, so you can run it by hand to see where the disk went.

The useful part is what the author learned the hard way. Deleting the session files directly leaves "Session not found on disk" entries, because the app keeps the Runs list in memory and writes records back. So deletion goes through the app's own delete_session tool, which is disabled inside routine runs and takes 25 sessions per call, each with its own approval card: 571 runs is 23 cards. The README also flags a setting worth checking even if you never install it: Claude Code deletes everything under ~/.claude older than cleanupPeriodDays, 30 by default, and it suggests 3650.

Reach for it if your desktop app's routine list has grown into the hundreds. Delete the afternoon you were going to spend clicking delete. Tradeoffs: macOS only in testing, with Windows and Linux paths written but untested; a few hundred runs is still a few dozen approval cards; it only handles routine sessions, not regular ones; and the Show HN post had 3 points.

e2e (Apache-2.0, TypeScript, 3,804 stars, 627 commits, 24 contributors, created July 22, 345 stars on GitHub trending today) is a test runner for web and mobile apps where a step can be a goal: agent.act('upgrade the workspace to the Pro plan'), followed by ordinary locator assertions in the same test. Web runs on Playwright's three browsers, iOS and Android through simulators, and a reporter posts results to the pull request. There is no default model and no shared key: you pick a subscription (npx e2e login), any AI SDK provider with tool calls, or a local OpenAI-compatible server. A decision-model executor can run bounded steps on Jev or Clef instead, where model output "never becomes a selector, a URL, or a key". Tests without agent steps need no model.

The replay cache is what keeps the model bill flat. An agent.act step is recorded only after a later check verifies the result, and the next run repeats its actions with no model call. The replay must start on the same route, find each control by role, name and context, and see at least one recorded change happen during the replay itself, so an outcome already on screen proves nothing. Dates, ids and counts that change between runs are ignored. If anything doesn't match, the agent takes over from the current screen, and the run summary counts replayed, handed-off and missed steps. Friday's pick IronBee Express had the same record-and-replay idea, but needed a hosted decision model for every live step.

Reach for it if you want goal-style end-to-end tests and want to choose, or host, the model they run on. Delete the hosted model key your agent tests needed just to click the same button every run. Tradeoffs: it is pre-1.0 and says APIs and config can change between minor releases; the CLI sends anonymous telemetry unless you set E2E_TELEMETRY_DISABLED=1; agent.assert always runs live, so assertion-heavy suites still pay per run; and it is the open-source front of TesterArmy's paid testing platform.

Also on the desk: Red Hat's benchmark of decision models against traditional guardrails (88 points on HN) is the outside evidence the last three weeks of Jev-alikes were missing. The headline overstates it: Jev had the best content-safety accuracy of nine guardrails, at 86.20%. But on prompt injection, a DeBERTa classifier scored 89.01% at 54.1 ms, against Jev's 86.35% at 348.1 ms.

Verification notes: star, commit, contributor, creation and release data are from the GitHub API on 2026-10-05; commit totals use the per_page=1 Link-header last-page count. The Thompson and Apple quotes are from the Stratechery post of 2026-10-05. Strata's architecture, speeds, requirements, size ratings and the Coder's SWE-bench figure are from its README and docs/MODELS.md; the paper citation, accuracy complaint, URL report, 60-versus-21 report and llama.cpp contribution rule are from HN item 49953495, whose title is the submitter's own 4090 result. RemoveMacAI's mechanism, commands and side effects are from its README; the disk figures, consent quote, install objection and mobileconfig suggestion are from HN item 49957116. SCM's models, presets and per-image timings are from its README; the 12,000-video question, the M1 report and the Tesseract exchange are from HN item 49952111. Routine Cleanup's buckets, the delete-tool limits and the cleanupPeriodDays advice are from its README; the 618 figure is from HN item 49946507. e2e's model options, replay rules and telemetry are from its README and docs/models.mdx, docs/cache.mdx and docs/decision-models.mdx. The Red Hat figures are from its 2026-10-02 article; HN points are from the Algolia API (items 49962857, 49953495, 49957116, 49952111, 49946507, 49933476). Pipeline health: today's pool held 80 rows, against the ~280 the compose prompt assumes. github:trending returned 12 rows, and Reddit returned zero again, 149 days since its last nonzero day on May 9.