I recounted these from the repository tree: 91 directories in plugins/, 202 agent files, 181 SKILL.md files, 105 command files. It matches what is claimed, which for a project this size is rare.

A plugin's shape shows in a single example:

The promise about context is stated separately: installing a plugin loads only its components into context, not the whole marketplace.

Installation

For Claude Code it is two lines, and the catalog is added by repository name:

After installation, commands are invoked with the plugin prefix, and /plugin lists everything available:

Plain words work too, and then the model picks the specialist:

Codex and Cursor install from committed registries that point at the same source directory:

Antigravity and OpenCode go through cloning and generating, because the built artifacts are not in the repository:

For those two the documentation is honest about what you give up: there is no one-step-from-URL install, only a local path after generation. That is named as the lean tradeoff.

Five Harnesses or Six

Here the project disagrees with itself, and it is worth knowing so you do not hunt for something that is not there.

The README says "one source-of-truth, five harnesses", the build command is labelled make generate-all # all five, and the table right under that phrase lists six: Claude Code, Codex CLI, Cursor, OpenCode, Antigravity CLI and Copilot. In the separate supported-harnesses table in the harness documentation there are five - Copilot did not make it in, even though it has its own output paths, its own install command, and its own test in the suite.

One more divergence from the recent past: Antigravity CLI has taken Gemini CLI's place. Anything written before August 2026 names Gemini - looking for it in the current repository is pointless.

An Honest Table of Losses

The most valuable thing here is not the number of agents but a document that usually does not exist: a capability matrix spelling out, per harness, what gets lost.

Practical consequences follow. If your skill references TodoWrite, on four harnesses out of six it leans on something that is not there. If the exact tool allowlist for an agent matters to you, Cursor and Codex will not preserve it. And Codex's 8-kilobyte cap on a skill body is why long skills have to be split.

The recommendation about the context file is stated separately: about 150 lines or 500 tokens, and it is the same across all harnesses. The principle is phrased as: the context file is a table of contents, not an encyclopedia.

What the Adapter Handles Itself

The author, meanwhile, does not need to know each harness's rules. The adapters resolve incompatibilities mechanically, and that too is laid out in a table.

So you write model aliases once as words - fable, opus, sonnet, haiku - and in each harness they turn into whatever it understands.

Model Tiers

A model is not assigned to an agent arbitrarily but by the tier of the work.

Authoring Rules

If you plan to write your own plugins, four principles are taken from OpenAI's harness-engineering post and stated firmly.

The context file is a table of contents, not an encyclopedia: up to 150 lines, with detail in docs/ or in a skill's references/. The repository is the system of record: what is not in plugins/ or docs/ the agent cannot see, and the places where knowledge does not count are listed explicitly - not chat threads, not documents, not notes. Enforce invariants, not implementation: frontmatter shape, file naming and trigger-phrase conventions are checked mechanically, while style within those bounds is yours. And a preference for boring tech: Markdown, YAML frontmatter and small Python adapters, with no templating engines and no custom markup languages.

A separate note on the trigger phrase in a description. It is mandatory, and specific wordings are recognised: "Use when", "Use this skill when", "Use PROACTIVELY when", "Use after", "Trigger when", "Auto-loads when". Without one, the MISSING_TRIGGER lint fires. That phrase is exactly what the model uses to decide whether to invoke your skill.

One more rule from the same family, useful outside this project: talk about actions, not tools. The models behind Codex do not have a Read, Edit, Bash vocabulary - they pick the native tool from the action you describe.

Quality Evaluation

The repository carries its own tool for evaluating plugins and skills, built in three layers.

The static layer is a deterministic structural analysis, claimed at under two seconds and free. The LLM-judge layer is semantic evaluation across four dimensions, around thirty seconds on Haiku and Sonnet. The Monte Carlo layer is statistical reliability through fifty to a hundred simulated runs, two to five minutes.

About those numbers, honestly: the run count for Monte Carlo is stated, but what the first two layers were measured on, how many times, and on what hardware is not written down anywhere. These are order-of-magnitude estimates, not a reproducible measurement.

Two External Entries

The catalog holds 93 plugins, but two of them do not live in the repository and are pulled in from other people's.

The first is Pensyve, external memory for the agent. It also has its own integrations for Codex, Cursor, OpenCode and Copilot, though not Antigravity.

The second is HOL Guard, security. And here the handling is worth calling out as exemplary: the external payload is pinned to a specific commit, its tool versions are pinned, and installation requires the user's explicit approval. It is stated separately that the catalog entry is only a discovery surface for Claude Code: the plugin does not enter the generated registries of the other harnesses. And it says outright that when protection is explicitly requested, the external program can change the harness's settings and hooks - but that is done by its own executable, not by files from this repository.

What to Know in Advance

The project has no version. Zero releases in the repository, no tags either. Each plugin is versioned separately in its own plugin.json - python-development, for instance, sits at 1.2.3. The internal evaluation tool has version 0.1.0, which must not be taken as the version of the whole catalog.

Generated artifacts are never edited by hand. The requirement is stated plainly: never hand-edit generated files. Before committing, the build for all harnesses has to be run, and CI checks for registry drift. The large generated skill trees do not enter the repository - only small registries pointing at the source do.

Installing everything is a bad idea. Users used to run into a warning about cumulative agent descriptions reaching four hundred thousand tokens against fifteen thousand allowed, after which Claude Code's commands became unusable. The current documentation promises that only the installed plugin is loaded into context, but the advice to install two or three plugins for your domain still stands.

Not everything in the catalog is about code. Among the plugins there are SEO, content marketing, sales automation, HR and legal compliance, deck creation, quantitative trading and game development. The catalog is noticeably broader than you would expect from a set of agents for a programmer.

There are three check commands, and knowing about them is more useful than knowing about a hundred agents: build all harnesses, verify structure, and hunt for drift and dead links.

First-Encounter Mistakes

Trying to install an individual skill. You cannot: the plugin installs as a whole.

Selecting skills by hand. Do not; they activate themselves when a description matches the task.

Installing everything. Agent descriptions accumulate in context, and that has already made things unusable.

Looking for Gemini CLI. Antigravity CLI replaced it, and old instructions will not work.

Expecting a one-command URL install for Antigravity or OpenCode. Their route is clone and generate, and the authors call that the lean tradeoff.

Referencing TodoWrite in a portable skill. Of six harnesses, two have it.

Writing a skill over eight kilobytes and expecting Codex to read all of it. It truncates at load; the adapter splits the body into a separate file.

Editing generated files. The build will overwrite them, and CI will catch the drift.

Forgetting the trigger phrase in the description. Without it the skill will not be invoked, and the lint will say so.

To put it in one sentence: the value of this repository is not its two hundred and two agents, but that one piece of Markdown you write works in six harnesses - and that there is a table stating honestly what each of them will lose.

Sources

This article was checked against the wshobson/agents repository (branch main) on 25 August 2026: the README, ARCHITECTURE.md, the Makefile, docs/usage.md, docs/harnesses.md, docs/authoring.md, docs/architecture.md, the .claude-plugin/marketplace.json registry and the plugins/ directory tree. The plugin, agent, skill and command counts were recounted from the repository tree and matched the claimed figures. Stars and forks were taken the same day. The project moves fast and publishes no releases, so the catalog's contents change without any version number changing.

• GitHub: wshobson/agents
• Installing and invoking
• The cross-harness capability matrix
• Plugin authoring rules
• Architecture
• The plugin catalog
• The agent catalog
• The skill catalog
• The quality evaluation tool
• Verification against real harnesses
• Pensyve, external memory