QuantSkills: a directory for quant Skills and Agents that AI runtimes can install

What QuantSkills actually is, and who the directory is for

QuantSkills is a navigator repository for the quantskills GitHub organization. It is not a quant library. The README describes it as a community directory of quant Skills and Agents that are discoverable, installable, verifiable and shareable, initiated by PandaAI and explicitly not an official PandaAI certification or investment advice. The repository you land on holds a README, an English README, a site directory, a scripts directory, a data directory, docs, tests and an assets folder. That layout is a catalog generator, not a backtesting engine.

The audience is narrower than the phrase "quant skills" suggests. It is aimed at people who already have an agent runtime, such as a code assistant or an agent framework, and who want to attach quant capabilities to it: a data connector, a factor generator, a data quality auditor, a risk monitor. The README's own framing is that both people and AI Agents should be able to find, install, review and combine these capabilities. If you are looking for a package that computes a factor for you after a pip install, this is the wrong entry point. It is an index of entry points.

One consequence is worth stating plainly. A directory can only be as trustworthy as the metadata its contributors supply, and the README leans on that: contributors are asked to describe data, assumptions, parameters, limits and risk boundaries before submitting. The catalog snapshot in the README is generated, carries a sha256 hash, and is dated 2026-09-01. It records 214 assets across 10 categories, with 1 published endpoint.

How the catalog is organized: ten categories and a workflow map

The README's catalog block lists the categories with asset counts, which is the fastest way to understand the project's shape. Category 01, data interfaces and warehouses, has 7 assets. Category 02, factor research tooling, has 44. Category 03, market and instrument analysis, has 44. Category 04, risk monitoring and alerting, has 22. Category 05, backtesting and trading tools, has 25. Category 06, research models and reproduction, has 30. Category 07, research validation and quality tools, has 12. Category 08, news search and knowledge analysis, has 10. Category 09, quant agents and automation, has 14. Category 10, infrastructure and templates, has 6.

Inside each category, assets are grouped by sub-topic and rendered as tables with columns for the project, a bilingual summary, a primary stage, inputs, outputs, interface status and a screenshot. That table is the real product. It is what lets you filter 214 entries down to the handful that match a pipeline stage.

A second view, the workflow map, arranges the same assets by pipeline stage rather than by topic. It defines five groups: data foundation (data-ingestion, data-quality), research signal (feature-engineering, factor-generation, factor-screening, modeling), portfolio validation (portfolio-construction, backtesting, evaluation, risk), monitoring and trading (monitoring, execution, reporting), and orchestration. This is the more useful of the two views if you are assembling an agent pipeline, because it tells you which stage is thin. Data foundation has 2 stages and 7 assets; research signal has 4 stages and 44-plus assets. The imbalance is visible in the numbers, and it matches how open quant projects usually accumulate: factor ideas are cheap to publish, clean point-in-time data is not.

The install path: clone, build, serve

The README does not give install instructions. What it does give is the repository layout and a package.json, and those are enough to reconstruct the intended workflow. The package.json marks the project private, sets the module type to ESM, and defines exactly two scripts: build runs node scripts/build.mjs, and test runs node --test tests/*.test.mjs. There is no start script, no dependency list and no published package name, so there is nothing to install from a registry.

That means the practical first use is to clone the navigator, run its build, and inspect the generated output. The commands below follow the two scripts in package.json verbatim.

git clone https://github.com/quantskills/quantskills.git
cd quantskills
npm run build

After the build, the README points readers at the interactive catalog at www.quantskills.ai, and the repository contains a site directory that the build script presumably populates. The README does not document what artifacts scripts/build.mjs emits, so treat the output as something to inspect rather than something to depend on.

The test script is the other half of the loop, and it is worth running before you change anything, because the catalog is generated from data and a broken generator silently produces a wrong index.

npm test

This runs Node's built-in test runner over tests/*.test.mjs. There are no third-party test dependencies in package.json, which suggests the tests exercise the build output rather than a running service. The README does not state what the tests assert.

Where the directory stops: metadata, not code review

The strongest limitation is stated by the project itself. The README says the catalog snapshot only displays public metadata and does not represent a quality endorsement, a return promise or a guarantee of production readiness. Read that against the interface status column, which marks most entries as pending maintainer review with no public endpoint, and exactly one asset as published. A directory of 214 assets where one has a published endpoint is a map of intentions as much as a map of software.

The second limitation is structural. Because QuantSkills indexes other repositories rather than vendoring them, every quality property you care about lives one hop away. The bilingual summary tells you a skill computes a limit-up tracking factor from dragon-tiger list seat history; it does not tell you whether the seat data is survivorship-free, whether the factor is computed point-in-time, or whether the backtest in the linked repository uses the same universe as the live signal. The README asks contributors to declare data, assumptions, parameters, limits and risk boundaries, but a declaration is not a verification, and nothing in the repository indicates the organization runs one.

There is also a scope mismatch worth naming. Several entries are explicitly described as research-layer candidate finders rather than trading strategies, and one factor entry says the weights can be retrained with ICIR plus shrinkage and includes a rolling IC gate. That is a research artifact with real modelling choices baked in. Treating it as a signal to trade is a category error the project warns against but cannot prevent.

How it compares with running your own research stack

The obvious alternative is a single integrated backtesting framework, where data loading, factor computation, portfolio construction and evaluation live in one codebase with one API. The difference in approach is not quality, it is coupling. An integrated framework gives you a consistent execution model and one place to debug; it also forces every strategy through the same abstractions and gives you nothing for the stages it does not cover, such as point-in-time fundamental reconstruction or futures roll auditing.

QuantSkills inverts that. It assumes your agent runtime is the integration layer, and each capability is a separately maintained package with its own interface status. You get coverage of stages a single framework usually omits: category 07 alone lists 12 assets for look-ahead and data leakage, survivorship bias, walk-forward and out-of-sample testing, signal stability, prediction calibration and workflow auditing. You pay for that coverage with heterogeneity. Two skills in adjacent categories may assume different universes, different date conventions and different output schemas, and the directory will not reconcile them for you.

A second alternative is simply reading the source of the handful of quant repositories you already trust. That is cheaper for one or two capabilities. QuantSkills becomes worth the hop only when you are assembling many stages and need to know what exists before you know what to search for.

Maintenance, licence and upgrade cost

The last push to this repository was on 2026-09-01, the same date as the catalog snapshot, so the generated index and the repository are in step. The repository is not archived. Beyond that, the project does not support claims about release cadence: no releases were retrieved, and the README does not describe a versioning or deprecation policy for indexed assets.

The licence is listed as unknown. That matters more here than in a normal library, because QuantSkills is a directory that links to other repositories, and each linked skill carries its own licence. The README asks contributors to describe data sources, which in quant work often means vendor data with redistribution terms attached. Nothing published states how licence compatibility across linked assets is checked, so the practical position is that you verify the licence of each skill you actually install, separately from this repository.

Upgrade cost is low for the navigator itself: it is private, has no dependencies in package.json, and its two scripts are a build and a test. The cost sits in the indexed assets. Because the directory does not pin versions of the repositories it links to, a skill that worked when you adopted it can change underneath you. If you build a pipeline from these assets, record the commit hash of each skill repository you pull in, since the catalog will not do it for you.

Editorial conclusion

Adopt QuantSkills if you are building an agent workflow and want a browsable index of quant capabilities with declared inputs, outputs and risk boundaries; skip it if you want a single production strategy package or a library with a stable API. Before relying on anything you find there, open the individual skill repository, check whether its interface status is listed as published or still pending maintainer review, and read the data, assumption and risk notes the contribution rules require.

Frequently asked questions

Official sources

1. Issues
2. quantskills/quantskills on GitHub
3. README