The Fetch: Week 31, 2026
Docs that behave, CRMs with homework, less robotic Chinese copy, light-beam file transfers, and wild C-based local inference
Anydoc: Office Files Finally Behave
github.com/firecrawl/anydoc | License: MIT
The Motion: Markdown Without the Document Drama
Anydoc turns Word, PowerPoint, Excel, EPUB, RTF, CSV, and PDF into clean Markdown with suspiciously little friction. The interesting part is its shared document model, which gives every format one consistent output instead of the usual converter roulette. That means tables, footnotes, headings, and speaker notes survive the trip in a way LLM pipelines actually want. Stars are showing up fast because this hits a very real pain point right now: teams are feeding messy legacy docs into agents, and brittle parsing is suddenly the bottleneck. Content-based format detection and sub-5ms speed make it feel ready for production already.
The Wave: The Ingestion Layer Agents Needed
This has a real shot at becoming the default document-to-Markdown layer for anyone building AI workflows, search pipelines, or internal knowledge tooling. Node.js and Python bindings widen the funnel, and the WebAssembly demo is a smart flex because local conversion matters when docs are sensitive. Honestly, the appeal is simple: one tool, lots of formats, no OCR bill for text-based PDFs. The next move is doubling down on trust with more public evals, edge-case fixtures, and format-level quality dashboards. That would make Anydoc feel inevitable for developers who need document ingestion they can actually reason about.
Stars: 4,537 | Language: Rust
Crm: Sales Software for Agent Believers
github.com/trycompai/crm | License: MIT
The Motion: The CRM That Actually Does Homework
This is not another CRM with a chatbot taped on. The pitch is sharper than that. Agentic-first CRM treats the database like storage for a durable research worker that keeps running even after the tab closes. The interesting part is the split architecture: the API just logs events, then the agent picks up queued work, researches contacts and companies, schedules follow-ups, and writes back only when evidence is strong enough. That evidence-first rule is why people are starring it now. It feels like one of the first repos to take “AI CRM” seriously without turning customer data into guesswork.
The Wave: Internal Tools Just Got a Lot Smarter
This could become catnip for startup ops teams, founder-led sales, and anyone tired of manually cleaning pipeline data that should have updated itself days ago. Features like the Agent tab, optional outside sources, and a sandbox with deny-all egress make it feel unusually thought through for such a fresh repo. Honestly, the ambition here is bigger than CRM. It hints at a pattern for internal software where agents own the busywork and humans settle edge cases. The next move that would make this unstoppable is tighter onboarding for self-hosters, because the faster teams can trust and deploy it, the faster this thing spreads.
Stars: 6,078 | Language: TypeScript
Human Writing: Chinese AI Copy Gets a Pulse
github.com/KKKKhazix/human-writing | License: MIT
The Motion: Anti Model Voice, Pro Actual Person
This repo is a writing skill for making AI-generated Chinese sound like someone specific said it, not a polished blob with no fingerprints. The interesting part is that Human Writing does not just ban obvious AI phrases. It targets the moves behind them, like fake reversals, empty buildup, canned lyrical flourishes, and repetitive paragraph drift. It also ships a check_prose script and a lite distilled version for chat windows, which makes it feel instantly usable. Stars are popping because Chinese AI writing has a sameness problem, and this attacks it with sharp, practical rules.
The Wave: Style Guides Become Agent Skills
This feels bigger than a prompt pack. Human Writing is really a portable editorial system for Chinese AI output, with rules for reality writing, fiction, formats, and revision baked in. That makes it interesting for agent builders, indie writers, content teams, and anyone tired of reading text that is technically fluent but spiritually dead. Honestly, this could become the default finishing layer for Chinese-language AI publishing. The next move that would make this unstoppable is stronger before-and-after examples across more genres, so people can see the voice shift immediately and trust it faster.
Stars: 1,001 | Language: Python
Decimen Optical Transfer: Files Through Pure Light
github.com/bashalarmistalt/decimen-optical-transfer | License: MIT
The Motion: QR Codes That Actually Move Bytes
Decimen Optical Transfer turns a screen and camera into a legit file link. The core trick is fountain-coded QR transfer, which means the sender blasts an endless stream of animated QR frames and the receiver only needs enough good frames to rebuild the file. No pairing, no Bluetooth drama, no shared network, no install. That gap is real, especially for airgapped devices, locked-down environments, and quick cross-device handoffs. People are starring it now because it feels like a browser demo from an alternate timeline, except it already works offline, preserves filenames, and verifies transfers with SHA-256.
The Wave: Airgapped Sharing Gets Weirdly Practical
The interesting part is how fast this could become the default answer for “how does this file get over there without touching a network?” Security folks, field teams, hardware hackers, and anyone moving files between hostile or isolated devices should be paying attention. Honestly, the no-app browser delivery is what makes this spreadable. It lowers the curiosity threshold to zero. The next move that would make this unstoppable is sharper guidance around privacy expectations and ideal transfer setups, because this thing is strongest when users instantly understand that no network does not mean encrypted. Very early, very clever, very sticky.
Stars: 4,676 | Language: TypeScript
Kimi K3 In C: Absurdly Real Local Inference
github.com/FareedKhan-dev/kimi-k3-in-c | License: Apache-2.0
The Motion: A Trillion Parameter Flex That Boots
Kimi K3 In C runs inference for a 2.78T parameter model on a single CPU with as little as 8.24 GB RAM, using portable C99 and literally zero framework baggage. The hook is not just CPU inference. It is the repo’s streaming trunk, MXFP4 experts, and byte-identical outputs across wildly different memory budgets. That means the same model answer on a laptop or a fat server, just slower or faster. People are starring it now because it turns frontier-scale model serving into a systems problem instead of a GPU shopping spree.
The Wave: Systems People Just Found Catnip
This has immediate pull for local AI tinkerers, inference nerds, and anyone obsessed with squeezing impossible things through ordinary hardware. The interesting part is how much trust the repo earns with its test gates, measurements, and docs that explain exactly where every byte goes. That transparency matters when the claim sounds fake on first read. The next move is making setup around the 1.56 TB checkpoint feel smoother, because the engine is already compelling and better onboarding would make it way easier for curious builders to actually cross the line from star to run.
Stars: 2,525 | Language: C