# AI Hacker Daily — 2026-08-13 Three model launches repriced tokens today.

"Why use many token when few token do trick." A Claude Code skill that instructs the model to answer in clipped, article-free prose, plus — since Monday — a local proxy and a Go compression engine underneath it. It plugs into Claude Code, Codex, Gemini, Cursor, Windsurf, Cline and thirty-odd other agents, exposes five MCP tools, and ships TypeScript and Python SDKs. The claim in the repo description is 65% of tokens cut. It has 97,905 stars, 5,649 forks and 243 watchers, which is not a ratio an organically-followed repo produces, and it crossed our own pool on July 2 at which point we did not cover it. Today it is trending again at 263 stars a day and launching on Product Hunt, four months after the original Show HN took 904 points. Here is what happened when people checked. Max Taylor ran 24 prompts across six categories through five arms — 120 scored responses, judged by a separate Claude against per-prompt rubrics for required facts, required terms and dangerous wrong claims. Baseline output: 636 tokens. Prepending the words "be brief." to the prompt: 419 tokens, a 34% cut. Caveman lite: 401. Caveman full: 404. Caveman ultra: **449** — worse than the two-word prompt. Quality across every arm landed within 1.5 points (0.985 baseline, 0.985 "be brief.", 0.975 full) with zero safety violations. His conclusion: "the value isn't compression. It's structure." In July, JetBrains ran their own test and got 8–10% savings with no measured quality degradation. And the honest part is that caveman's own documentation does not really dispute any of it: the README states the skill adds ~1–1.5k input tokens per turn and that "on already-terse workloads they can go net-negative," and HONEST-NUMBERS.md goes further, disclosing that Copilot's per-request billing makes the tool useless by construction ("A shorter answer is the same request"), and that one Cursor A/B measured **4.3M tokens with caveman versus 1M without, and double the wall-clock time**. It publishes a break-even rule that tells you to turn it off. The best comment on the JetBrains thread explains the whole gap in one line: "the bulk of my token usage is not what I type in — it's that data the AI is processing for me." Reach for it if you want consistent, structurally terse agent output and will A/B it against your provider's actual usage page, which is the only test the maintainer himself endorses. Delete the assumption that a compression tool's headline number describes your bill. Tradeoffs: the 65% figure is output tokens only on the vendor's own 10 tasks, ranging 22–87%, and output tokens are the cheap end of most agent workloads; the counterfactual 33.2% input-token figure comes from 54 runs across 18 test cases, published as a number rather than as raw runs; the license is split — MIT for the skill, CLI, SDKs and graders, BSL-1.1 for the engine, proxy, MCP binary and browser driver, converting to Apache-2.0 on 2030-06-21, with third-party hosted or embedded use requiring a commercial license; Taylor's benchmark is single-run per arm-prompt pair and he conceded in-thread that the variance findings may be noise; JetBrains tested at reasoning-low only; the v2.0.0 engine shipped Monday and the npm CLI was published two days ago, so the thing you install today is not the thing anyone benchmarked; and 476 open issues sit against those 243 watchers.

Dosu shipped a local-first reader for the session logs your coding agents are already writing. Point it at `~/.claude/projects` and `~/.codex`, and it parses the `.jsonl` transcripts into a SQLite archive and serves a dashboard on `127.0.0.1:3000` breaking spend down by token count, estimated cost, context-window occupancy, activity, tool calls, MCP servers and files touched. One command: `npx @dosu/decant@latest`. Apache-2.0, TypeScript, 23 stars. It is in this slate because it is the instrument for the argument the pick above lost. If the majority of your token bill is context the agent pulled in rather than prose either of you wrote, then compressing the prose is optimizing the wrong term, and the only way to know your own ratio is to read your own transcripts. Decant is the first tool we have seen aimed squarely at that question rather than at the total — every usage dashboard in this category tells you *how many*; this one is built to tell you *where*. The obvious objection is that it ships without answering the question for itself: the README has a screenshot of the dashboard and no findings, no sample distribution, not one number about where tokens typically go. For a tool whose entire premise is that the breakdown is surprising, publishing one real breakdown would have done more than the screenshot does. The privacy posture is stated plainly and is the right one — "it makes no outbound network calls at runtime, and your transcripts never leave your machine" — which is worth noting given that Dosu's actual business is a cloud product. Reach for it before you buy any tool that promises to cut your token spend, so you know which term you are cutting. Delete the guess about where your context window is going, and any dashboard that reports only a total. Tradeoffs: 23 stars, one fork, zero watchers, and a Show HN that drew 10 points and **zero comments**, so there has been no outside scrutiny of the parser at all; it reads two harnesses and breaks silently if either changes its on-disk layout, which is the same unfixed exposure we flagged on Inventory back on August 3; costs are estimated, not billed, so it will not reconcile against an invoice; v0.4.0 was published to npm the evening before the Show HN, and the package is three weeks old; and there is no published methodology for how it attributes a token to a category, which for this tool is the whole product.

A private AI gateway that sits between your organization and every provider you use, so that a request carries an identity before it leaves the building. Project-scoped API keys with team ownership, per-project quotas and concurrency limits, OAuth/OIDC with RBAC and audit trails, and usage logs attributed to user, project, team, model and cost center. It fronts OpenAI, Azure OpenAI, Anthropic, Gemini, DeepSeek, Qwen and local models with 150-plus provider templates, speaking both OpenAI-compatible endpoints and the Anthropic Messages API. Apache-2.0, Go, SQLite for a single instance and PostgreSQL when you need replicas, deployable as a systemd unit or Docker Compose. It is maintained by astaxie — the author of beego, which is a name worth knowing before you read 1,035 stars as unproven. The reason it made the slate over a dozen other gateways is in `benchmarks/`. There is a committed `budget.json` that reads: minimum success rate 99.9%, minimum throughput ratio 0.9, maximum mean latency regression 15%, maximum p99 regression 20%. Alongside it sit checked-in baseline files — smoke, streaming, responses and local-failover, each as both JSON and Markdown — plus the scripts that run them. That is a gateway committing, in CI, to a numeric ceiling on what it costs you to run everything through a gateway. Every product in this category asks you to accept an extra network hop in exchange for governance and essentially none of them tell you what the hop costs; this one publishes the number it refuses to regress past and fails its own build if it does. It is the same disclosure standard Ante set for agent benchmarks on Monday, applied to infrastructure, and it is the cheapest thing any tool in this category could do. Reach for it when more than one person is spending money against your provider accounts and nobody can say who spent it. Delete the shared API key in the team password manager, and the spreadsheet where somebody reconciles provider invoices to projects by hand. Tradeoffs: it is two months old with exactly two tagged releases, v0.4.0 and v0.5.0 a week ago; the budget file constrains overhead but no absolute latency numbers are published, so you learn what it promises not to get worse than, not what it is; its one Hacker News appearance in July took 2 points and no comments, so there is no independent operational report anywhere; the repo lives under a personal account while the docs site sits under a `thinkinai-labs` org, and the docs URL says `tokenhome` rather than TokenHub, which is a seam worth resolving before you make it the chokepoint for every model call in your company; 36 open issues against 6 watchers; and the quotas are per project and per concurrency, which throttles rate rather than capping spend.

Usage-based billing infrastructure for developers, self-hosted or cloud: real-time metering of arbitrary usage events, prepaid and promotional credit balances, automatic top-ups at a threshold, credit expiry, and feature entitlements defined as toggles, metered limits or config values tied to a plan. AGPL-3.0, Go, 4,300 stars, 493 forks, on its 23rd point release of the v2.1 line and pushed this morning. It is the rung where your token spend stops being your cost and becomes someone else's price — the layer every AI product that sells credits ends up writing badly, in-house, twice. It is also where this ladder runs out, and it is worth being precise about why. The README says you "can enforce usage limits in your application by checking with Flexprice." That is an entitlement API, not an enforcement rail: the platform tells your code the customer is out of credits, and your code is what decides whether the request goes through. Which means that after four rungs of increasingly good accounting — a compressor that measures its own overhead, a reader that attributes tokens to a cause, a gateway that attributes requests to a person and caps its own latency — the thing that actually halts spend is still a conditional somebody remembered to write. We have been keeping a watch open since July 27 for the first tool to publish how a hard agent-spend ceiling is enforced rather than metered, through Orite and hotcell and Paritok. This is the closest anything has come, it is open source, and it still does not close it. Nothing in this category stops. It only counts. Reach for it when you are about to hand-roll credits, metering and plan entitlements into your own product for the second time. Delete the bespoke usage table, the cron job that expires credits, and the Stripe webhook spaghetti between them. Tradeoffs: enforcement is your application's job and the docs do not pretend otherwise, so a bug in your check is an unbounded bill; roughly 99% of the code is AGPLv3 with the Enterprise Edition held back in `ee/` and `internal/ee/` — both directories verified present — and AGPL is a real constraint if you embed it in a hosted product; there are no published throughput, event-volume or scale numbers anywhere in the repo, which for metering infrastructure is the number that matters most; 4,300 stars against 9 watchers; and its own Show HN attempts through the spring never cleared single-digit points, so the traction here is GitHub's, not a community's.

Everything above measures a bill. This one deletes one. Woxi is a Wolfram Language interpreter written in Rust, AGPL-3.0, and it has been under development for two years — 8,279 commits — before landing on Hacker News yesterday and taking 296 points, the day's largest Show HN. It runs as a CLI, a Jupyter and JupyterLite kernel, a Python package, an npm package, a WASM module in the browser, and Woxi Studio, a native notebook editor built with iced that opens `.nb` files and exports to `.ipynb`, Markdown, LaTeX, Typst and PDF. `cargo install woxi`, `pip install woxi`, or `npm install woxi-wasm`. The README points at a `functions.csv` for implementation status and does not total it, so we did. The file lists 6,295 Wolfram Language functions. 2,985 are marked implemented — **47.4%** — with 1,804 explicitly marked as won't-do (external services, GUI infrastructure, notebook UI, control systems) and the rest unmarked. That headline number undersells it badly, because the file also carries a usage rank. Of the 500 most-commonly-used functions in the language, **485 are implemented**; of the top 100, 91. So the honest description is not "half of Mathematica" but "essentially all of the Wolfram you would reach for in a script, and none of the long tail." That is exactly the right shape for the stated goal — "a subset of the Wolfram Language so that it can be used for CLI scripting and notebooks" — and it is a distinction no headline, including the project's own, currently makes. Reach for it when a Wolfram dependency is the only thing keeping a script off CI, a container, or a colleague's laptop. Delete a per-seat Mathematica license from the machines that only ever run scripts. Tradeoffs: 47.4% overall coverage is real and the long tail is where symbolic mathematics actually lives, so anything leaning on advanced symbolic manipulation, probability distributions or mesh generation will hit a wall fast; the won't-do list is permanent by design, not a roadmap; AGPL-3.0 is a genuine constraint for commercial embedding, which is a notable inversion given what it replaces; 804 stars, 26 forks and 7 watchers after two years means a very small number of people have leaned on this in anger; there is no published conformance suite against Mathematica's own output, so "implemented" is the author's assessment per function and nobody has independently graded a single one; and it is the deliberate off-theme pick today — the best pure artifact in the pool, and the only one that is not an AI tool.

**Also worth knowing.** **Delta** (564 points) is Zed's new multiplayer environment for coding with agents — a Rust application compiled to WebAssembly, built on the DeltaDB that crossed our pool a week ago, replicating the conversation and the worktree together so you can comment on code at any point in its history rather than only at commits, and pull a terminal Claude Code session into a shared thread. It is the day's biggest tool story and it is **private beta with no license stated**, so it fails "installable today" and sits here instead. **Someone is running mass vulnerability scans while spoofing ClaudeBot and other AI crawlers** (278 points) is the attribution problem from the opposite end of the wire: today's picks make your agent traffic legible on your side, and this is the reminder that nothing makes the label on inbound agent traffic true. **cpa-usage-keeper** (MIT, 1,138 stars, pushed this morning) is a genuinely nice SQLite-backed usage dashboard whose substrate is CLIProxyAPI — i.e. routing API-shaped traffic through CLI subscription credentials — so it is excluded on the same grounds as the proxying tools we dropped in July, not on quality. **GreenTree** (Apache-2.0, 6 stars, created yesterday) caches test results by git tree rather than commit SHA, so a dirty working tree that matches a previously-verified tree skips the run — the same "stop paying for work you already did" instinct one layer over, and far too new to pick. **Trunchbull** (6 points) runs models against Harbor-authored benchmarks from the browser, which is the natural completion of Monday's Ante watch about whether anyone would actually pull those raw runs; its site would not load at compose time, so it is named rather than assessed. **kepano/obsidian-skills** (MIT, 45,307 stars, 3,279 forks) trended today at 252 stars a day on a repo whose last commit was in June — worth a note that GitHub trending now surfaces things nobody has touched in two months. Dropped per rubric: DeepSeek V4 Pro (945), Qwen3.8-2.4T (638), Grok 4.6 (560) and Codex Desktop for Linux (173), all launches; the SQLite WAL-reset bug from Tailscale (1,057) and Antithesis's own writeup of the same 16-year-old bug from the other side (118), which together are the best engineering reading of the day and are not tools; and "I took the calculator away from 50 LLMs and graded the arithmetic" (7 points), which is the right kind of experiment sitting in a repo with no license and no stars. The pool was 90 — hn:front 28, ProductHunt 24, hn:show 19, github:trending 19 across five feeds. Reddit produced nothing for the fifteenth consecutive week.