GitHub AI Weekly · Issue 03 (Aug 29, 2026) · AI Beyond

Top 10 projects

1. 01

   freestylefly/awesome-gpt-image-2

   MultimodalJavaScript

   This week+12,877
   

   Prompt as Code — a GPT-Image-2 prompt engine and template library, with 530+ reverse-engineered cases and 20+ templates packaged as skills you can load into a coding agent.

   Stars

   24,238

   Forks

   2,396

   Open issues

   17

   Last commit

   2026-08-28

   The pace is the story: four separate commits this week added cases 521 through 544, roughly a batch every other day, and the companion gallery lets you filter by style and copy a full prompt in one click.

   If you already have a way to call GPT-Image-2 and would rather not start from an empty prompt box, editing one of these cases beats ten rounds of guessing.

   That said, this is a prompt library, not an image tool — you bring your own model access. The README gives more space to sponsors than to content, with several API relay vendors mixed in among the writeups, so read it with that in mind. Prompts are also version-bound: when the model updates, older cases may stop reproducing.

2. 02

   tt-a1i/archify

   AI codingJavaScript

   This week+11,099
   

   An agent skill for Cursor, Claude Code, Codex CLI and OpenCode: the agent emits typed JSON, and a local renderer compiles it deterministically into self-contained HTML/SVG architecture, workflow, sequence and data-flow diagrams.

   Stars

   27,415

   Forks

   1,739

   Open issues

   61

   Last commit

   2026-08-28

   Honestly, there is barely a feature commit in the repo this week — what changed was site language switching, a star-history chart and where the viewer's navigation dock sits, and the last tagged release, v2.15.0, is still dated Aug 14. Those eleven thousand stars are word of mouth from the previous fortnight cashing in at once.

   Anyone already writing technical docs with these coding agents who wants the architecture diagram to be checkable against the code will at least stop hand-writing Mermaid.

   The catch is that mainline currently sits at 2.16.0-dev, so features land on the dev version first; output is self-contained HTML/SVG, and real polishing still means opening another tool; and the README and product page are more finished than the documentation, so don't judge maturity from the showcase renders.

   Deep dives

   Last issue this slot belonged to cathrynlavery/diagram-design. This week that project dropped off the board entirely and archify took its place. Same job — get a coding agent to draw an architecture diagram worth looking at — two winners in two weeks. The demand is clearly firmer than any single repository serving it.

   The approach is worth a paragraph. Archify does not let the model emit SVG or Mermaid directly. The agent produces a typed JSON intermediate representation, and a local Node program validates and compiles it. That moves where errors surface: if the model invents a call that does not exist, validation can reject it, whereas a hand-drawn diagram from a model looks identical whether the edges are real or not. It also diffs two snapshots into Before / Delta / After, listing added, removed, changed and rerouted nodes one by one — that is a code-review artifact, not a slide.

   The cold water: no release shipped this week, so what grew was attention, not capability, and there is a full round of changes sitting between the tagged version and dev. Before putting it in front of a team, run it once on a mid-sized repository and see whether it recognises your layering. Drawing something pretty and drawing something correct are different problems.

3. 03

   openai/codex

   AI codingRust

   This week+9,109
   

   Lightweight coding agent that runs in your terminal, written in Rust, with editor and desktop builds alongside the CLI.

   Stars

   119,566

   Forks

   18,254

   Open issues

   14,287

   Last commit

   2026-08-29

   A dozen-plus 0.151.0 alphas went out this week, almost all of them subagent and MCP plumbing: cleaning up the bundled browser when a subagent stops, running executor hooks even for interrupted turns, per-tool MCP output limits, preserved context baselines across nested agent forks, cloud task credentials restricted to trusted origins.

   If you already pay for a ChatGPT plan and live in a terminal rather than an IDE, one sign-in gets you working against that plan's quota.

   But there are more than fourteen thousand open issues, and several alphas can land in a single day, so tracking mainline means accepting the occasional broken build; the recommended sign-in path is tied to a ChatGPT plan, and using an API key instead takes separate setup.

4. 04

   MadsLorentzen/ai-job-search

   AI agentsPython

   This week+4,828
   

   The job search that runs on your machine: an application framework built on Claude Code where /setup, /scrape, /apply and /interview cover profiling, sourcing, fit scoring, tailored CVs and cover letters, and interview prep. Fork it and own it.

   Stars

   37,735

   Forks

   12,768

   Open issues

   14

   Last commit

   2026-08-27

   This week's commits are all about applicant tracking systems: extract PDF text with pypdf and fall back to Poppler, so you can verify the generated LaTeX CV is actually readable to a screener, plus caching company research so the second role at the same employer doesn't trigger another lookup.

   If you are applying right now, have Claude Code available, and are willing to move your CV onto the repo's LaTeX template, this fits the way you already work.

   The constraints are concrete: the portal scraping targets Danish job boards (Jobindex, Jobnet, Akademikernes Jobbank), so another market means rewriting that layer against the same pattern; the whole pipeline runs on Claude Code and the token bill for each round is yours; and whether to tell employers that AI produced your materials is a call you have to make — the author's own choice was to say so upfront.

   Deep dives

   The convincing part of this repository is not the 37,000 stars, it is the account on the first screen of the README. The author is a geophysicist by training whose position was cut in late 2025. He wrote this to run his own search, sent 69 tailored applications through the same /scrape, /apply and /interview flow, got 20 first interviews, signed one contract, and started as an AI engineer in June 2026. He also notes that he told every employer how the materials were produced, and that it usually opened a technical conversation rather than counting against him.

   Structurally it reads as a disciplined workflow rather than an agent toy. Fit evaluation scores a posting and recommends, application materials go through a drafter-then-reviewer pass, and the CV is generated through LaTeX so formatting stays stable — this week's pypdf extraction exists to prove the result survives a screener. The 12,000 forks tell you people are using it the way it was designed: fork it, fill in your own profile, keep it.

   Know what you are picking up. The reusable core is the self-profiling and application pipeline; the sourcing half was written for one country's job boards and mostly needs rewriting elsewhere. What it saves is the Sunday spent writing cover letters, not the decision about where to apply.

5. 05

   Alishahryar1/free-claude-code

   AI infrastructurePython

   This week+4,769
   

   Use Claude Code, Codex, Pi, OpenCode, Cline, Aider and five more coding agents against one catalog assembled from 50 providers' free tiers, with automatic fallback to the next model when retries run out — no restarting the turn.

   Stars

   51,201

   Forks

   8,245

   Open issues

   359

   Last commit

   2026-08-28

   This week it added Aider as the tenth harness, switched Muse Code to a native Windows install, and fixed two long-standing translation-layer bugs: images being dropped across protocols, and cache-hit accounting not adding up after a translated response.

   If you want to try several coding agents side by side without paying a separate subscription for each, one install lets you switch between them against the same catalog.

   But the word free is defined by each provider, the rules change without notice, and the project itself says it removes an integration once a provider stops allowing it; most of the 359 open issues sit in that translation layer, where debugging spans several hops; and handing several vendors' keys to one local service is a risk you should price yourself.

   Deep dives

   Two projects in this week's top ten are doing the same thing: pooling free tiers behind one endpoint so coding agents can draw on them. This one is fifth, freellmapi is tenth. That lane did not exist in the last issue; this week it holds two slots.

   It does two things a plain relay does not. The first is protocol translation: Claude Code speaks Anthropic's protocol, Codex speaks OpenAI's, and the model underneath may be neither, so something has to translate both ways — this week's fixes for images lost in translation and cache hits not reconciling live in exactly that layer. The second is token thrift: an optional RTK filter strips the bulk of useless terminal output, which the project claims cuts terminal-output tokens by up to 90%, alongside several optimisations aimed at quota probes and command prefixes. It also makes a point of following provider terms, dropping integrations that are no longer allowed, and states plainly at the top of the README that it is not affiliated with Anthropic.

   Whether it is worth installing comes down to how much you trust the convenience to last. Free tiers are not commitments; a combination that works today can be gone next month. As a low-cost sandbox for trying agents it fits well. As a cost-saving plan for production work it is a bad bet — for that, one honest paid quota is less trouble.

6. 06

   rohitg00/ai-engineering-from-scratch

   AI infrastructurePython

   This week+3,263
   

   A free, MIT-licensed AI engineering curriculum built around writing things by hand: 511 lessons across 20 phases and roughly 329 hours, in Python, TypeScript, Rust and Julia, where every lesson ships an artifact — a prompt, a skill, an agent or an MCP server.

   Stars

   50,618

   Forks

   8,780

   Open issues

   98

   Last commit

   2026-08-23

   The last substantive change was the Aug 23 batch that added an agent-skills track and a stateless MCP curriculum, and collapsed the MCP route to a single canonical path. Nothing landed in the repo this week; the stars are momentum from people still passing the course around.

   If you want to work through the ladder from linear algebra to agents and MCP in order, and will actually build the artifact each lesson asks for, the sequencing is the value here.

   Bear in mind that 511 lessons and roughly 329 hours assume you finish, and most people won't get halfway; English is canonical while the other languages are machine-translated lesson pages on a separate branch; and with no commits this week, you'll want to check the pace yourself before committing months to it.

7. 07

   VoltAgent/awesome-agent-skills

   AI agents

   This week+2,372
   

   A curated collection of agent skills from official dev teams — Anthropic, Google Labs, Vercel, Stripe, Cloudflare, Hugging Face, Figma and more — alongside community submissions, with the compatible clients listed per entry: Claude Code, Codex, Gemini CLI, Cursor and others.

   Stars

   33,063

   Forks

   3,484

   Open issues

   18

   Last commit

   2026-08-26

   A dozen-odd community submissions merged this week — a spec writer, a Chinese de-AI-ifier, a multi-source search skill among them — and the badge count has moved from the 1000+ written in the README body to 1497+.

   If you run Claude Code, Codex or Cursor and have no appetite for writing your own skill, browsing the vendor sections for the official ones is the fastest path.

   It is still a list of links, so whether a skill installs cleanly is the upstream repo's problem; the community section is populated by pull request, and this week's merges range from serious tooling to a dark-psychology skill pack, so don't take the list on faith; and since the repo holds no code, anything that breaks goes back to the source project.

8. 08

   tinyhumansai/openhuman

   AI agentsRust

   This week+2,353
   

   A local-first personal AI: your data compressed into scored Markdown memory trees in SQLite, mirrored as an Obsidian vault you can open and edit, with agent orchestration and deep research running on top.

   Stars

   38,724

   Forks

   3,802

   Open issues

   325

   Last commit

   2026-08-27

   The through-line this week was consolidating memory scoring into its own event bus and pinning every module's tinymemory dependency to 1.13.2 to stop submodule and registry versions drifting apart — foundation work on the memory layer rather than new surface.

   If you want the AI's memory of your life to live on your own machine in files you can open in Obsidian, rather than in another cloud assistant, this is the one currently doing that.

   It ships labelled Early Beta by its own authors; GPL-3.0 is a hard constraint for anyone hoping to embed it in a closed-source product; figures like 100+ OAuth integrations, 5,000+ MCP servers and 90,000+ skills are the project's own claims; and with 325 open issues and desktop versions moving fast, expect rough edges.

9. 09

   anthropics/claude-plugins-community

   AI codingPython

   This week+2,207
   

   The community plugin marketplace for Claude Cowork and Claude Code as a read-only mirror, synced nightly from Anthropic's review pipeline; plugins appear only after passing automated security scanning and approval.

   Stars

   2,585

   Forks

   213

   Open issues

   42

   Last commit

   2026-08-25

   It has 2,585 stars in total and 2,207 of them arrived this week — this is essentially the week it got discovered. The reason is mundane: installing community plugins requires an address to hand `claude plugin marketplace add`, and this is that address.

   If you want community plugins for Claude Code or Cowork but would rather someone ran a security scan before you install them, this is the vetted door.

   Only, it is a read-only mirror — pull requests opened here are closed automatically and submissions go through the official form; the catalog syncs nightly, so what you see is not necessarily the current state; and the repo holds no plugin code, so implementations live upstream and entries disappear when upstream does.

10. 10

    tashfeenahmed/freellmapi

    AI infrastructureTypeScript

    This week+2,162
    

    Free tiers from 34 providers plus any custom OpenAI-compatible endpoint behind a single /v1 API: keys stored encrypted, a router that picks an available model and fails over when one is rate-limited, and per-key usage tracking to stay under every free-tier cap.

    Stars

    21,614

    Forks

    3,061

    Open issues

    66

    Last commit

    2026-08-28

    Five releases went out on Aug 28 alone, v0.8.6 through v0.9.0: the Mac desktop build is now signed and notarised (the DMG itself, not just the app inside), an estimated usage frame fills in when an upstream omits usage so quota accounting doesn't silently drift, and an opt-in Fetch Relay outbound transport landed.

    If you want a dozen providers' free allowances to behave like one /v1 endpoint and are willing to run the router yourself, that is exactly what this does.

    But the rules for those free tiers belong to the providers, and the project scopes itself to personal experimentation; the model catalog comes from a signed feed where free installs get a monthly snapshot — new models reach you 30 days late unless you pay $19 a year; and the README carries a dense layer of promotional links worth reading past.