Directories / Open sourceThe builder’s edit

Ideas for your next build

Open source.
Worth opening.

The AI projects worth a closer look. What they do, what you could build, and what to watch out for. Less directory digging. More getting somewhere.

Try a checkable build Explore the shortlist ↗

Selected for usefulness. Not ranked by lifetime stars.

01 Real things to build02 Primary sources, linked03 Trade-offs included

The useful part of the changelog

What changed. Why care.

Three source-linked changes with a practical interpretation.
Editorial assessments are pinned to a specific release, not generated from a version number.

vLLM · 0.32.0

Teams operating model APIs

New structured decision endpoints and expanded KV cache options create new integration and memory trade-offs to evaluate.

Your next move

Benchmark a representative workload before upgrading. New serving options are not universal speed guarantees.

Read the source release ↗

Docling · 2.140.0

Builders processing mixed document and audio sources

Remote transcription and richer AsciiDoc parsing widen the input paths available to a document pipeline.

Your next move

Test one example of your actual input format. Check where remote transcription sends data.

Read the source release ↗

The shortlist

Six projects.
Six reasons to pay attention.

An opinionated starting point, not an exhaustive ranking.
Sources checked 19 Oct 2026. Not installation tests.

Agent harness01

Deep Agents

Give an agent more than a chat loop.

Planning, files, context management and delegation in one extensible harness. A useful starting point when a task takes more than one tool call.

Try building

A research assistant that gathers evidence and writes a report.

Know first Constrain tool permissions and require approval for consequential actions.

Choose this when…

Choose it: You need planning, files and delegated steps in an agent, without assembling each piece yourself.
Skip it: A single prompt or deterministic script solves the job. A harness adds state and failure modes.
Consider instead: LangGraph for a custom execution graph ↗
Skills & hardware: Python or TypeScript; tool permissions and testing. API models need network access; local models need their own inference hardware.
Setup: 30–60 min for a small prototype (estimate). Not a measured installation time.

Coding agent02

OpenCode

Choose the coding agent. Keep your model choice.

An open-source coding agent for working in a real repository. Worth exploring if you want to choose your provider rather than commit to one model ecosystem.

Try building

A small feature in an existing app, with tests and a reviewable diff.

Know first Review generated changes. Model usage and hosted services can still cost money.

Choose this when…

Choose it: You want a coding agent in your repository with a choice of model providers.
Skip it: Your team needs a particular editor integration or enterprise control this project does not supply.
Consider instead: Aider for a different terminal-first workflow ↗
Skills & hardware: Git, code review and a project test command. Ordinary development machine for hosted models; local model requirements vary.
Setup: 15–30 min with a configured provider (estimate). Not a measured installation time.

Document processing03

Docling

Your PDFs deserve better than copy and paste.

Convert documents into structured representations for search, retrieval and downstream AI. Start here before building another chatbot on messy source files.

Try building

A searchable document library with structure and page references.

Know first Check extraction quality on your own layouts, tables and scans.

Choose this when…

Choose it: Layout, tables and multiple document formats matter to your retrieval pipeline.
Skip it: You only need text from a simple text-based PDF. Start with a smaller parser before adding model assets.
Consider instead: pypdf for a text-PDF baseline ↗
Skills & hardware: Python; extraction-quality checks. Requirements depend on the pipeline and models. First run can download assets.
Setup: 20–45 min including initial downloads (estimate). Not a measured installation time.

Web data04

Crawl4AI

Turn a website into something AI can actually read.

A self-hostable crawler that turns web pages into Markdown and structured data. Useful when the hard part of your AI workflow is getting usable source material.

Try building

A research pipeline that collects public pages with their source URLs.

Know first Respect site terms, robots policies and rate limits. Do not bypass access controls.

Choose this when…

Choose it: You need repeatable web collection with browser-backed extraction and Markdown output.
Skip it: An official API or a direct download already gives you the data. Prefer that simpler, permitted path.
Consider instead: An official API, or Playwright for custom browser logic ↗
Skills & hardware: Python; selectors, site policies and rate limits. A supported browser runtime; parallel crawling uses extra memory.
Setup: 20–45 min including browser setup (estimate). Not a measured installation time.

Local AI & training05

Unsloth

Make open models your own.

Run and fine-tune supported models with a local interface and training tools. A practical route from using a model to adapting one to a specific task.

Try building

A fine-tuning experiment with a small, carefully checked training set.

Know first This release is a beta. Check GPU memory, model licences and training data rights.

Choose this when…

Choose it: You have a measurable task, rights-cleared training data and a supported model to adapt.
Skip it: You have not tried retrieval, better prompts or evaluation first. Fine-tuning is not a way to invent missing facts.
Consider instead: Ollama when you only need to run a supported model ↗
Skills & hardware: Dataset preparation, evaluation and Python or the project UI. Model, quantisation and training method determine GPU and memory needs.
Setup: 45–90 min for environment and a small experiment (estimate). Not a measured installation time.

Model serving06

vLLM

Go from a model on disk to a model people can use.

Serve supported open models through an API, with batching and memory-conscious inference. Explore it when your project needs a shared model endpoint.

Try building

An internal model API, benchmarked against your actual workload.

Know first You own hardware, authentication, monitoring and capacity planning.

Choose this when…

Choose it: Several users or services need an inference endpoint, and you can operate the infrastructure.
Skip it: You need a first local experiment or do not want to own capacity, access controls and monitoring.
Consider instead: Ollama for a simpler local starting point ↗
Skills & hardware: Model serving, deployment, authentication and load testing. Supported accelerator and enough memory for the chosen model and workload.
Setup: 45–90 min with provisioned hardware (estimate). Not a measured installation time.

Leave with something working

Small builds. Visible proof.

Runnable scripts, sample inputs and explicit checks.
Document and research samples were executed. Local AI has client contract tests, not a model benchmark.

Actual sample run

A PDF becomes checkable data

Extract three invoice fields with a page number and exact supporting text. Missing or conflicting values stay unresolved instead of becoming guesses.

10–15 min setup estimateOpen the mini-build →Actual upstream run

A research brief with receipts

Collect two pinned release records into a compact evidence brief. Preserve the release URL and source excerpt before anyone writes a summary.

5–10 min setup estimateOpen the mini-build →Contract tested; inference not run

A local model with a known-answer check

Send a synthetic extraction task to your own loopback Ollama server, then reject any response that does not match the expected answer.

15–45 min setup estimate, excluding model downloadOpen the mini-build →

Go one step beyond browsing

One document.
A useful first build.

Try Docling on a short, non-sensitive PDF. Get Markdown you can inspect before feeding it into a search or AI workflow.

01Choose a sample with a heading, a paragraph and a table. Save it as sample.pdf.
02Install Docling in an isolated Python environment. Save this example as convert.py.
03Run it, then compare the Markdown with your original. Check reading order, numbers and table structure.

Official quickstart & prerequisites ↗

convert.py

from docling.document_converter import DocumentConverter

converted = DocumentConverter().convert("sample.pdf")
print(converted.document.export_to_markdown())

In your terminal

python -m pip install docling
python convert.py > sample.md

Your check: a readable sample.md that agrees with the PDF. Missing or scrambled content is a reason to investigate, not to trust the output.

Source-reviewed recipe, not tested on your hardware. Initial setup may download model assets. Follow the project’s installation and privacy guidance.

Follow the work

Fresh from the repositories.

The latest release of each of the 500 most-starred open-source projects in the directory, newest first.
635 of them publish GitHub releases. Read every night, last checked 10 Nov 2026. Editorial review is separate.

10 Nov 2026