Skip to main content
← Back to the Journal
PRODUCT · Open source·May 2026·12 min read

Twenty open-source AI tools — GitHub links before the star count.

You leave a creator breakdown with a list of names; the next step is not "save for later" — it is clone, read the README, and separate "interesting demo" from "runs on customer money or personal data."

This piece continues the open-source dev tools series: links upfront because that is the practical ask, then the Nuqta question — what do you tell legal if they sit next to you tomorrow?

The batch — copy/paste GitHub links.

Finance and market data — reference code is not a license to trade.

anthropics/financial-services ships analyst-style agent patterns; internal sign-off still precedes production use on portfolios or customer files [2].

OpenBB is the workstation layer — Bloomberg/LSEG contracts remain separate engineering and procurement work. Tie spend to GCC procurement TOR discipline, not hype graphs [3].

Agents, skills, training harnesses — more repos, wider blast radius.

lobehub/lobehub and hello-agents lean into multi-agent orchestration; flutter/skills proves skills are not web-only [4][5][6].

addyosmani/agent-skills, mattpocock/skills, and garden-skills snapshot conventions — healthy if someone owns merges, dangerous if they rot [19][20][21]. Every skill wired to an agent reopens enterprise MCP boundaries: who owns keys and outbound tools?

jcode is an experimentation harness; governance sits with whoever approves merges to main, not whoever runs the coolest prompt [7].

Local research, faster decoding, durable workflows.

local-deep-research mixes search backends with LLMs — replay its README benchmarks on your corpora, not only the public web [8]. dflash chases speculative-decoding latency wins that matter when you already fund GPUs for local SLM economics [9].

sdk-typescript is how you model multi-day workflows; budget operators, not only npm upgrades, before agents depend on it [10].

The link saves ten minutes of searching — it does not replace a processing register. If nobody owns liability after the README, you installed curiosity, not tooling.

Red lane — trading, stealth browsers, OSINT.

AI-Trader advertises autonomous trading — legal reviews first, screenshots second [11]. CloakBrowser is powerful for authorised security testing; outside that charter it is evasion tooling [12]. Maigret aggregates public footprints — fine inside an approved investigation, toxic for casual employee surveillance under PDPL-style obligations [13].

Desk, notes, niche glue.

obsidian-copilot brings models into vaults — watch agent mode exports and sub-processor paths [14]. writer-computer keeps prose on disk when SaaS residency is a non-starter [15].

oh-my-design emits design-system markdown — judge it on token alignment, not logo count [16]. wasabi is VBA networking glue; narrow, but honest about where "enterprise integration" really starts [17]. ds4 is pedagogy metal, not your next microservice dependency [18].

Lane radar — one glance.

FIG. 1 — MAY 2026 GITHUB RADAR BY GOVERNANCE LANE

Before a repo hits the approved list.

  • Classify data touching the tool — personal, regulated market data, or operational-only sets PDPL review depth [1].
  • Write RACI for CVE patching and for license interpretation (MIT + bad transitive deps is still legal work).
  • Run an exit drill: revoke API keys, delete agent transcripts, clear scratch storage.
  • Multi-agent stacks need knowledge layers — read agents vs RAG-first before stacking skills on empty corpora.

Closing.

You have the same links we wish every roundup included first — the rest is operational honesty: who signs, where data lives, when the experiment stops. If one of these repos collides with your compliance reality, send the war story — the Nuqta Journal is built from desks, not thumbnails alone.

Frequently asked questions.

  • Does OSS exempt PDPL in Oman? No — law follows processing, not repo licenses [1].
  • Why isolate AI-Trader? Automated capital decisions hit market rules READMEs never cover [11].
  • Does Local Deep Research replace internal RAG work? It accelerates harnessing, not retrieval quality on your documents — read the RAG guide.
  • Are GitHub skills a substitute for security policy? No — files in a repo do not replace scoped tools, secrets, and logs.
  • Tiny team? Cap yourself at three repos this quarter — governance beats checklist length.

Sources.

[1] Sultanate of Oman — Personal Data Protection Law 2022 (Royal Decree 6/2022).

[2] Anthropic — anthropics/financial-services on GitHub.

[3] OpenBB — OpenBB-finance/OpenBB on GitHub.

[4] LobeHub — lobehub/lobehub on GitHub.

[5] Datawhale — datawhalechina/hello-agents on GitHub.

[6] Flutter — flutter/skills on GitHub.

[7] jcode — 1jehuang/jcode on GitHub.

[8] LearningCircuit — LearningCircuit/local-deep-research on GitHub.

[9] DFlash — z-lab/dflash on GitHub.

[10] Temporal — temporalio/sdk-typescript on GitHub.

[11] HKUDS — HKUDS/AI-Trader on GitHub.

[12] Cloak — CloakHQ/CloakBrowser on GitHub.

[13] Maigret — soxoj/maigret on GitHub.

[14] Obsidian Copilot — logancyang/obsidian-copilot on GitHub.

[15] Writer — joelbqz/writer-computer on GitHub.

[16] oh-my-design — kwakseongjae/oh-my-design on GitHub.

[17] Wasabi — uesleibros/wasabi on GitHub.

[18] ds4.c — antirez/ds4 on GitHub.

[19] Addy Osmani — addyosmani/agent-skills on GitHub.

[20] Matt Pocock — mattpocock/skills on GitHub.

[21] Garden Skills — ConardLi/garden-skills on GitHub.

[22] Nuqta — May 2026 rewrite emphasising inline GitHub links.

Related posts

Share this article

← Back to the JournalNuqta · Journal