# AI Hacker Daily — 2026-08-27 Today's tools are sorted by how much of the company is still inside them.

A BSD-3 Go CLI and library from Tailscale — `go install github.com/tailscale/tailcat/cmd/tailcat@latest`, or `nix run github:tailscale/tailcat` — that does what netcat does over what Tailscale does: one side runs `tailcat` and prints a token, the other side runs `tailcat <token>`, and stdin on one machine is stdout on the other through a WireGuard tunnel that bootstraps over a DERP relay and then hole-punches to a direct UDP path "when possible (usually!)." No account, no root, no routing-table or DNS changes; it is gVisor's netstack terminating TCP inside the process. `--serve=8080,8443` forwards local ports; `--serve=no-auth-ssh` is an SSH server with no auth; `tailcat socks <token> curl ...` runs a command through a SOCKS5 proxy over the tunnel; `--serve=exit-node` is an exit node; a token in a DNS TXT record means `tailcat ssh example.com` reaches a machine with no open inbound ports "where WireGuard authenticates the client before the SSH server ever sees a packet." It was open-sourced at TailscaleUp this month, hit the front page yesterday at 586 points, and has 1,309 stars, 22 forks and 10 watchers. Brad Fitzpatrick has made 47 commits in the last seven days; overnight, between 01:07 and 01:41 UTC, he published a WebAssembly demo to GitHub Pages that sends files and text from a browser to the CLI. The company inside it is one JSON file. The default relay map at `tailcat.dev/derpmap.json` lists four regions — New York, San Francisco, Frankfurt, Tokyo — and the README's Stability section is the most candid on today's slate: "The public rate-limited Tailcat DERP relays have no uptime SLAs or throughput targets, and we may revoke access to them at any time, for any reason." The next section is titled "Contact Sales?" The fix is a flag: `--derpmap-url` points at a relay you run from `cmd/derper` in the main repo, and Fitzpatrick's thread answer to the lock-in complaint was the whole design in a sentence: "There's no vendor lock-in here and no payment or account required. If Tailscale as a company fails, tailcat keeps working if you run your own DERP server." Two details from the history. The tool began "in September 2023 as 'derpcat', written on a long flight while catching up on bad movies," bitrotted inside a fork of the tailscale.com repo, and was unforked in March — which is why a repository created in October 2024 carries 10,396 commits back to a 2020 "Initial commit." And the key model is the part to read before sharing a token: ephemeral keys by default, an address that dies with the process; `tailcat genkey` for a stable one, after which "anyone you've *ever* shared that address with can connect to any future server using that key, unless you restrict clients with `--allow`." tptacek's read is the right frame: "It's Magic Wormhole but for generalized connectivity, not just file transfer." The thread's other question — what stops malware using free, hard-to-block relays for command and control — got "rate-limited" as the answer, which is a throttle, not a policy. Reach for it when the alternative is ngrok, a bastion host, or explaining Tailscale onboarding to someone who needs one port for one afternoon. Delete the port-forward and the `ssh -R` tunnel through a VPS. Tradeoffs: "no API or CLI stability promises" — flags, output and the wire format "may all change"; there are zero releases, so installation means a Go toolchain or Nix; the token is a bearer credential unless you pin clients with `--allow`; browser peers are DERP-relayed only until WebRTC lands (#4); the public relays are Tailscale's goodwill, which the CEO has held "for ~6.5 years" and the README says can end any day; and the four open issues filed this morning — idempotent `Close`, a process-wide `netns` side effect, a PTY hang on slow shells — are what a two-week-old public tool looks like.

An MIT Rust CLI and daemon plus a Chromium extension — `curl -fsSL .../install.sh | sh`, `irm .../install.ps1 | iex` on Windows, the extension from the Chrome Web Store or Edge Add-ons, then `bsk install-skill` — that gives any shell-capable agent a separate, visible Agent Window inside your real browser profile, with your cookies and logins, while you keep using the other windows. The agent shells out to `bsk`; `bsk` talks to a local daemon over IPC; the daemon talks to the extension over a WebSocket bound to loopback; the extension drives the Agent Window. "The agent never talks to the browser directly." Nine harnesses get a one-step skill install — Cursor, Claude Code, Codex, OpenClaw, Pi, Hermes, DeepSeek Harness via an npm plugin, and Tencent's own CodeBuddy and WorkBuddy — and the README's claim is "no lock-in to a specific model, Agent framework, or harness." Created June 22; 1,395 stars, 110 forks, 7 watchers; `cli-v0.1.11` and `ext-v0.1.7` both shipped yesterday, and a "dsh browser tools parity" PR merged at 08:29 UTC this morning. The store listing, under the publisher `wxg.prc.cpg`, reports 10,000 users and a privacy declaration that it will "not collect or use your data." Read the two machine-facing files, because the security model is in them and not in the README. `docs/architecture.md`, "Security (v1)," is four bullets: loopback-only daemon and WebSocket, an extension-origin allow-list at the upgrade, "no credential storage in bsk — cookies stay in the user's browser profile," and `evaluate` restricted to Agent Window tabs. That last line is the load-bearing one, because the Agent Window is *in* the profile — the feature is that the agent is already signed in everywhere you are. The `SKILL.md` then does what Hands did on Monday: it puts the policy in markdown. Under "When NOT to use": "**Credential harvesting** — never run `bsk evaluate` on banking, SSO, or password-manager pages to extract tokens, cookies, or secrets." Most of the file is about stopping — "No post-success work," "at most one extra `bsk snapshot`," "When blocked, pause — do not brute-force," `bsk session stop` "REQUIRED when done (even on error paths)" — a skill whose main instruction to the agent is to leave. Set that beside yesterday's OpenAI timeline, where agents told to find a flag decided, in their own chain of thought, that "this is arguably unauthorized... Yet goal solution," and the sentence in `SKILL.md` is the fence. The recommended install is the same shape: paste "Set up browser-skill on this machine by following https://raw.githubusercontent.com/Tencent/BrowserSkill/main/AGENT_INSTALL.md" into your agent, and the agent runs the `curl | sh` for you. Reach for it when the agent needs a site you are already logged into — a deployed page to smoke-test, a dashboard to read, a form behind SSO — and the alternative is a Playwright profile you re-authenticate every morning or handing the agent the whole desktop. It replaces the headless browser with a test account, and it is one rung up from Hands on the sandbox-distance ladder: a window instead of your mouse. Tradeoffs: the isolation is a window, not a profile, so every cookie you have is one `bsk evaluate` away and the guard is a sentence the agent is asked to obey; the extension is the one component you get from a store rather than from source, and it updated yesterday; Chrome and Edge only, Firefox "planned"; the daemon auto-starts on any `bsk` call; the DeepSeek Harness path injects native `browser_*` tools that bypass the CLI shape; and there is no benchmark of any kind — the README has a video.

An Apache-2.0 skill from JetBrains — `/plugin marketplace add JetBrains/go-modern-guidelines` then `/plugin install modern-go-guidelines@goland-claude-marketplace` in Claude Code; `codex plugin add ...` in Codex; a Cursor marketplace; `npx skills add JetBrains/go-modern-guidelines` for the rest — whose README opens with "guidelines for code agents that help them write modern Go code" and does not pretend a human is the reader. The diagnosis is two sentences: "Training data lag. Models don't know about features added after their training cutoff," and "Frequency bias. Even for features the model knows, it often picks older patterns" — more `for i := 0; i < n; i++` in the corpus than `for i := range n`. Fifty-four guidelines in `internal/guidelines/guidelines.json` cover Go 1.0 through 1.27 "including everything targeted by the `modernize` analyzer"; the skill installs a small Go CLI into `~/.cache` on first use, reads the target version from `go.mod`, and the agent calls `list --file-path` before editing and `explain <id>` before skipping a rule. 1,856 stars, 53 forks, 18 watchers, and 314 stars today, the second-highest climb on the all-languages feed. Created last November, redesigned August 11, `encoding/json/v2` guidelines added August 19 in v1.1.1. The `SKILL.md` sets the trust level, and it is total: "Treat returned guidelines as authoritative for modern Go style choices... If a guideline applies, follow it even when nearby code or repository convention uses an older pattern. Skip it only when it would not compile, would change behavior, or clearly does not match." Yesterday two people checked whether the rules meet their own bar. `mxssl` filed PR #13 at 12:19 UTC: the `new_expression` example used `new(30)` for a `*time.Duration` field, "which does not compile"; all 193 indented example lines used spaces, so "the style reference handed to a code-generating agent" was demonstrating non-gofmt Go; and behavior-change caveats lived only in `explain`, while `SKILL.md` frames `explain` as a gate on *skipping*. At 15:54 `RO-29` filed #14 — "Eight defects: six examples that change behaviour, one wrong `since_version`, and wrong Go version detection. I measured each one with Go 1.27.0. Every number below comes from code that ran. None of it is an opinion" — and PR #20 four minutes later. The six are the kind that pass tests: rewriting `make`-plus-`range` to `maps.Clone` turns a nil source into a nil map and "the next write to the result panics"; `strings.CutLast` on a path with no separator swaps `dir` and `file`; `max` over a slice with a NaN returns NaN where the manual comparison returned 5; `defer stop()` after `context.AfterFunc` cancels the cleanup the goroutine used to run. The `maps_clone` details text presents the change as a feature — "It preserves nil maps" — and that sentence is still on `main` this morning; the last push is August 19 and both PRs are open with zero maintainer comments. JetBrains has spent this month as the third party that ran other people's numbers (ponytail, caveman, rtk); yesterday it got one. `FEATURES.md` had already said it: "Work in progress — inconsistencies may be present." Reach for it if your agents write Go and their output looks like 2019, which is the README's honest description of every model. Delete the "prefer modern Go" paragraph in `CLAUDE.md` and use a versioned list with a `since_version` on every rule. Tradeoffs: install the rules with #14 in mind and read `explain` for anything touching nil maps, `Cut`, `min`/`max` on floats, or `AfterFunc`; it needs a Go 1.25+ toolchain on `PATH` because the marketplace install runs `go install`; auto-update is off by default for third-party marketplaces, so the fixes, when merged, will not reach you unless you turn it on; and the product it markets is GoLand and Junie, which is fine, and is why the Claude Code and Cursor paths exist at all.

An MIT agent skill — `npx skills add tt-a1i/archify -g`, then "Use archify to map this repository's runtime architecture" — for Cursor, Claude Code, Codex and OpenCode, plus a DeepSeek Harness bundle since v2.15.0. The agent writes a typed JSON IR for one of five diagram types (architecture, workflow, sequence, data flow, lifecycle); a Node renderer compiles it deterministically into a self-contained HTML file with inline SVG, dark and light themes, PNG/SVG/WebM export and a 1200×630 share card; and nothing is delivered until `node bin/archify.mjs validate <type> <candidate.json> --quality showcase --json` passes. No telemetry; `preview` binds `127.0.0.1` on a random port and watches one file; "hosted sharing, and WYSIWYG editing are intentionally outside the current scope." It is the top repository on GitHub's all-languages trending today at 1,035 stars gained, 20,970 total, 1,374 forks, 91 watchers; created April 15; v2.15.0 shipped August 17 with 107 "provenance-backed" brand marks, digest-pinned logo fetches that "fail closed on drift," and a before/delta/after comparison of two validated snapshots for reviewing architecture changes in a pull request. The `SKILL.md` frontmatter credits the lineage — `based_on: Cocoon-AI/architecture-diagram-generator (MIT, v1.0)`. It first crossed our pool on July 3 via Product Hunt and has never been on the slate. Two things in the repository matter more than the star count. The first is `benchmarks/ordinary-model-floor`, which asks one narrow question — "can an ordinary coding agent produce an Archify diagram that is usable on attempt 1, without a human repairing the JSON?" — and calls itself "a delivery gate, not a model leaderboard." A run passes only if the semantics bind, the real CLI validates at showcase quality, and a named reviewer inspects the render; "a missing visual review is reported truthfully, never upgraded to a pass," and "reference fixtures are not benchmark evidence." The published result, July 26, three cheap models through Pi with a 30-minute limit and frozen first attempts: 10 of 15 usable — MiniMax M3 4/5, Qwen 3.7 Plus 3/5, DeepSeek V4 Flash 3/5. A third of first attempts fail the project's own gate, and the project published that. The second is the top of the README, which is now a sponsor table: an API reseller with a referral link and a 10%-off code, a memory-infrastructure company whose harness "supports Archify as a Skill," and "Want to sponsor Archify? Contact us by email." Between 04:41 and 06:11 UTC this morning the maintainer added a community link to a Chinese forum, removed it, and restored it, in three commits. That is the company inside a 21,000-star skill: not a vendor, a sponsor. Reach for it when the agent's answer to "how does this service work" is a paragraph and you want a file you can open, click, and diff against last month's. Delete the Mermaid-in-a-code-block habit — Archify will take Mermaid as input but is "not a general-purpose drawing editor or a Mermaid theme." Tradeoffs: "validated" means the JSON is well-formed and the geometry is clean, not that the topology is true — the instruction to "inspect repository evidence when the diagram must reflect real code" is addressed to the agent, and the benchmark's semantic check is against a manifest, not a codebase; a 67% first-pass rate on cheap models means a repair loop on the rest; the runtime is Node and the output is a house style in four presets you either like or do not; and 21,000 stars for a diagram skill is a fact about what people ask agents to draw, not about the diagrams.

A TypeScript code-intelligence engine — `npx gitnexus analyze` in a repository, `npx gitnexus setup` to wire the MCP server, `npm install -g gitnexus@latest` — that parses fourteen languages with Tree-sitter, resolves imports, calls and types across files, clusters the result into communities, traces execution flows, and indexes it all with BM25 plus embeddings under reciprocal rank fusion, so the agent's `query`, `context` and `impact` tools return callers, callees and blast radius with a confidence percentage instead of grep output. It writes an `AGENTS.md` and `CLAUDE.md` block into your repo unless you pass `--skip-agents-md`, installs skills into `.claude/skills/`, ships cosign-signed Docker images built only from version tags, and has a one-click Render blueprint at about $35 a month whose README says plainly that the access token "is the only control on this deploy" and that the proxy "strips `Origin` before forwarding, so the server's CSRF guard does nothing." 45,829 stars, 5,065 forks, 152 watchers; created August 2025; release candidates are tagged per merge, `v1.6.10-rc.225` at 08:43 UTC this morning. It relaunched on Product Hunt yesterday as "the open-source kernel for coding agents," #9 with 93 votes; it first crossed our pool on May 22 at 269 stars in a day. The license is PolyForm Noncommercial 1.0.0. GitHub's parser reports it as NOASSERTION, the README badge says it, and the Enterprise section says the rest: "fully managed SaaS or self-hosted deployment. Commercial use of the OSS version is also available with proper licensing," from Akon Labs. That is not an open-source license by anyone's definition, and the Product Hunt page says open-source in the tagline and the first line of the description. Second, the number. The maker's launch comment says coding agents run "51% cheaper with GitNexus connected" per a public benchmark. The figure does not appear in the README, in any of the four copies of the review skill, or in the benchmark directories. What the repository does contain is better than the number: an `eval/` harness that runs SWE-bench Lite against baseline bash tools versus GitNexus-enhanced tools across Haiku 4.5, Sonnet, Opus, MiniMax, GLM and DeepSeek, with no results published in the README, and a `bench/impact-pdg` harness whose own conclusion about its two impact engines is "they answer different questions; neither strictly dominates." The first line of the README is a warning that "GitNexus has NO official cryptocurrency, token, or coin," which is the kind of sentence a 45,000-star repository ends up needing. Reach for it if your agent spends its first twenty tool calls on `grep` and `cat` to learn what calls what, and the repo is big enough that the answer changes weekly. It replaces the hand-written architecture section of `CLAUDE.md` and the "read these five files first" ritual — for personal and noncommercial work. Tradeoffs: at a company you need a license before you index the monorepo, and the tagline will not tell you; indexing is memory-bound (the browser UI caps near 5,000 files, the server defaults to a 2 GB buffer pool); the tool inserts itself into your `AGENTS.md`; the cost claim lives on a launch page, not in the repository, and the harness that could substantiate it is there for you to run; and the `serve` mode has no authentication unless you set a public origin.

"Claude-English → plain English, as a service" — the repository's own subtitle. A Claude Code plugin (`/plugin marketplace add tenkenco/declaude`, `/plugin install declaude@tenken`, `/declaude:setup`) that registers a hook to rewrite every reply, an MCP server at `speak-english.tenken.co/mcp` that any client can add with OAuth and no pasted key, and a document translator. It strips "sycophantic openers, hollow superlatives and hedging filler" and runs Qwen2.5-14B-AWQ "on our own L4 GPUs, so your text never reaches a commercial AI provider," with prompt logging disabled at the model server. Free tier 100 rewrites and five documents a month; $5 a month for unlimited. The MIT repo has one star and was created August 12; the Show HN got seven points and four comments; the poster's reason for building it is the useful part: even with skills and a `CLAUDE.md`, "claude actually starts to remake the mistakes I told it not to make," so they moved the fix out of the prompt and into a second model. This is the third implementation of the same joke in five weeks, and the first to become a business. gvzdv's `claudish-to-english` is a local hook with a 17 GB Gemma; it had 1,976 stars when we ran it on August 21 and has 2,315 today. Vomit, our pick that day, was the Go CLI version. Declaude credits gvzdv's hook as its origin and moves the model off your machine. The 08-21 question from `bob1029` — "if we need to babysit 100% of its output with another vendor's model. Why not just use that other vendor's model for everything?" — now has a price attached: five dollars a month to have a 14-billion-parameter model on a rented L4 unwrite a frontier model's prose, on every turn, with a network round trip in the hook. Reach for it if you have already lost the `CLAUDE.md` argument about tone and you would rather pay than run a 17 GB model locally. It replaces the local hook and the manual rewrite. Tradeoffs: every reply transits a third party's server, and "processed in memory and discarded" is a promise, not a property you can check; the hook adds latency to every turn; a one-star repository two weeks old is the whole track record; and the underlying problem is upstream, where a model that a paying customer describes as "virtually intolerable" in a thread going "a bit viral on reddit" is the product being fixed at the edge.

**Also on the board.** *OpenExecutive* (652 points, 414 comments) is the AI CEO built by developers who, per the poster, were laid off to make room for AI — a FastAPI plus Next.js app where one "Executive" persona fronts eight Claude sub-agents over ChromaDB, `make dev`, your own Anthropic key, 1,014 stars. The `LICENSE` file is Apache-2.0 though GitHub's parser reports NOASSERTION (the license habit, resolving direction), and the `CLAUDE.md` is candid where the README is not: "Breaking caching = 10x cost increase," and "The internal agent architecture is NEVER exposed to the user." The thread's best line was about legal accountability — an AI cannot be liable, so a human stays — which is also why it is a chatbot with personas and not a CEO. • *acceptmarkdown.com* (145 points) proposes `Accept: text/markdown` content negotiation so agents get prose instead of DOM, with recipes for Nginx, Caddy, Rails, Next, Django and Cloudflare Workers; nobody in the thread or on the site says who built it. The useful artifact is the status page: seven agents already send the header — Claude Code, Copilot Chat and CLI, Cursor, Microsoft Copilot, OpenClaw, OpenCode — Codex fetches HTML and looks for a `<link rel="alternate">`, and eleven do not, including Claude.ai, Gemini, Cline, Windsurf and Zed. JimDabell: "Claude Code's been doing this for at least a year already." The thread's two objections stand: a second representation of one URL is a prompt-injection channel, and Akamai does not cache responses that carry `Vary`. • *Twitter Viewer* (465 points) is what fills Nitter's gap: Cloudflare-fronted, Hong Kong-registered, ads on every page, and a `.com` that Verisign can take on request; `zedeus/nitter` itself gained 63 stars today, the day after its cease-and-desist. • The fleet-view category from Monday got two more entries: *luvus* (AGPL-3.0, Rust, 472 stars, +97) is a tmux-shaped TUI that detects nineteen agents and shows status, tokens and cost per pane, installed by `curl | sh`; *tutti* (Apache-2.0, 3,494 stars) is a desktop app where Claude Code, Codex and Hermes share a workspace, with a "Tutti · VM" multi-user tier on a waitlist. • *open-slide* (MIT, 7,165 stars, +166) is Archify's shape for decks — "a slide framework built for agents," React on a 1920×1080 canvas, `/create-slide`. • *godogen* (MIT, 6,395 stars, +159) has Claude Code or Codex build a Godot, Bevy or Babylon game and judge it "from the running game... not from a clean compile" — it needs Google, xAI and Tripo3D keys and a 15–20-second proof video, and the last commit is July 26. • *TexLite* (AGPL-3.0, 28 stars) is a self-hosted Overleaf; yesterday's TeXbrain compiled in the browser, this one runs on your server. • *Mold* (140 points) is Rui Ueyama's arXiv paper on the linker you already use — "2.4–16.1x faster than lld," up to 112x over GNU ld — a paper about a product, which is the one kind of paper that passes. • *The Harness Is the Thing* (129 points, Scott Fryxell) is Sunday's harness theme as a personal essay: cheap models for the explore-and-work stages, frontier for plan and critique, "the fulcrum from which my expectations meet the LLM's capabilities." • Product Hunt: *Traccia* (#3, 114 votes) is a vendor-neutral agent control plane with an open Python and Node SDK on OpenTelemetry — the category is now six deep; *IQ Routing* (86 votes) is trajectory-aware model routing with no repo and no number. • Dropped per rubric: Nvidia–Hugging Face (1,206), AWS–DuckLabs (1,058), Stripe–Clerky (185) and Mechanical Turk (388) as news, all used above as the why-now; GLM-5.3-Flash (1,041 points, 525 comments) and Qwen3.8-Flash-Next as model launches; Nebula Sans (426) as a typeface; Risklytics as a brokerage. *Verification note: stars, watchers, forks, licenses, creation dates, release tags, issue and PR states and commit logs come from the GitHub API on 2026-08-27. Tailcat's relay list was read from `tailcat.dev/derpmap.json` (four regions); its history, key-management and Stability sections from the README at this morning's commit; the commit depth (10,396, back to 2020-01-31) and the March 6 "Unfork" issue from the API; Fitzpatrick's quotes are his own comments in the August 26 thread. BrowserSkill's security model is `docs/architecture.md` ("Security (v1)") and `skill/SKILL.md`; the store figures (10,000 users, v0.1.7, publisher `wxg.prc.cpg`) are the Chrome Web Store listing today. The JetBrains defects are issues #13–#20 and PRs #13 and #20 as filed on August 26, all still open or folded into #14; `internal/guidelines/guidelines.json` on `main` was checked this morning for the `maps_clone` "preserves nil maps" text (present) and counted at 54 entries; the last push is August 19. Archify's benchmark numbers are `benchmarks/ordinary-model-floor/results/2026-07-26-pi-three-models.json` (15 runs, 10 first-pass usable), the sponsor table and scope lines are the README, and the three LINUX DO commits are #133, #135 and #138. GitNexus's license is the `LICENSE` file (PolyForm Noncommercial 1.0.0); the "51%" figure was searched for in the README, all four `gitnexus-review` skill copies and the `bench/` and `eval/` trees and found only on the Product Hunt page. Declaude's model, hosting and pricing are its README; gvzdv's star counts are the API today and our August 21 edition. The OpenAI timeline is the company's own post and embedded incident timeline, read directly (May 12 first message-board entry; July 10 fourteen Hugging Face write tokens; July 11 HDF5 and RefJinja zero-days; July 12 multi-cluster access and four-region credentials; July 19 OpenAI Kubernetes admin); the Nvidia price is The Information's as cited in the thread, and the DuckLabs structure is the DuckLabs announcement as summarized in its thread. The machine-facing-docs check ran on all six picks: policy-stating on BrowserSkill (`SKILL.md` and `architecture.md` carry the security model the README omits), constitutive on the JetBrains repo (the `SKILL.md` sets the trust level the issues test), procedural-with-a-benchmark on Archify, sprawling on GitNexus (`AGENTS.md`, `CLAUDE.md`, `GUARDRAILS.md`, `SECURITY.md`, `RUNBOOK.md`, `DoD.md`), absent on Tailcat, and a plugin manifest only on Declaude. Seen-before SQL keyed on `fetched_at`: Archify first crossed July 3 (Product Hunt) and GitNexus May 22 (github:typescript); Tailcat, BrowserSkill, the JetBrains repo and Declaude are first-time; Tailscale itself was named in four prior notes (July 6, August 3, 6 and 13) and JetBrains in two (August 13 and 26, both as the party running someone else's benchmark). Today's pool was 88 — hn:front 30, hn:show 22, producthunt 21, github:trending 15 rows across five feeds (all 5, python 4, typescript 4, go 1, rust 1), and Reddit dark for the twenty-second consecutive week.*