This is all of the agentic tooling I use with Claude Code every day, kept in one repository and installable as a plugin marketplace. There are four plugins. mightymodels carries a ticket from the first investigation through review and cleanup. vibe-code builds and checks the skills, hooks and agents that everything else is made of. python-harness reviews Python the way I would. terraform-harness gives Claude Code the tools to read, check and change Terraform.
The repository runs on an "everything is a plugin" mindset, an idea I took from the DeepSeek
harness. The root owns the marketplace manifest and the shared quality gate. Every capability
ships as a plugin under plugins/, one directory each, with its own manifest, skills and agents.
terraform-harness is the exception: it lives in its own repository,
cc-terraform-harness, and its marketplace
entry points at a release tag there.
| Plugin | Version | What you get | Details |
|---|---|---|---|
| mightymodels | 1.0.0 | Takes a piece of work from the first question to a merged PR, with nineteen skills, twelve worker agents, seven hooks and a state server | README |
| vibe-code | 0.4.0 | Eleven skills for building and tuning Claude Code skills, subagents, hooks, rule files, MCP servers, loops and plugins, plus a CLI that validates what they write | README |
| python-harness | 0.2.0 | Eight Python skills, four agents, a ty language server, two hooks, a CLI and an MCP server | README |
| terraform-harness | 0.1.0 | Four Terraform agents, five skills, three hooks, the terraform-ls language server and an MCP server that checks modules, digests plans and looks up provider schemas | README |
No plugin depends on another. Install the ones you want.
Check this table before you install anything. Without uv, a plugin installs fine and then its
CLI, hooks or server will not start.
| You need | Needed by | Why |
|---|---|---|
| Claude Code | all four | They are Claude Code plugins. The vibe-code validators also need a claude plugin validate that accepts --json: 2.1.289 has the flag and 2.1.205 does not. |
uv on your PATH |
all four | The CLIs, hooks and MCP servers start through uv tool run, or uv run in terraform-harness. |
| Python | all four | Nothing to install by hand, because uv fetches it. mightymodels, python-harness and terraform-harness ask for 3.14 or newer, vibe-code for 3.12 or newer. |
| Linux, macOS or WSL | mightymodels, python-harness, terraform-harness | mightymodels starts through sh and bash scripts, the python-harness launcher does not run on native Windows, and the terraform-harness hooks start through a sh script. |
git |
mightymodels, python-harness, terraform-harness | mightymodels keeps a ticket's files in .mightymodels/ at the git toplevel, and its hooks look up the ticket for the checked-out branch. python-harness runs git to list the files a diff changed, and terraform-harness does the same when it traces a change's impact. |
terraform, tflint and terraform-ls on your PATH |
terraform-harness | The hooks and the MCP server run terraform and tflint, and terraform-ls is the language server. The session-start hook names any that are missing, and the checks that need a missing one are skipped. |
GitHub CLI (gh), logged in |
mightymodels | Issues, pull requests and CI watching. Without it, open-ticket keeps the issue body as a file and stick-the-landing stops after showing you the gh pr create command. |
jira CLI |
mightymodels, optional | Only when you pick Jira as a ticket's tracker. Without it you get a draft file. |
| Network access | all four | A package index, the first time uv builds each environment. The two python-harness documentation tools also read docs.python.org. |
Note
vibe-code is the only plugin with a Windows launcher, bin/vibe-code.cmd. It was tried by hand
on one Windows machine and no CI job covers it. The
vibe-code README has what was observed.
To work on the repository itself you also need Node 22 or newer, make, and the Claude Code CLI
at the version the gate was built against:
npm install -g @anthropic-ai/claude-code@2.1.285/plugin marketplace add rhawk117/rygentic-harness
/plugin install mightymodels@rygentic-harness
/plugin install vibe-code@rygentic-harness
/plugin install python-harness@rygentic-harness
/plugin install terraform-harness@rygentic-harness
The first line reads this repository's .claude-plugin/marketplace.json. Each line after it
installs one of the plugins that file lists. docs/claude-code.md covers
model ids, agent discovery and validation.
This is the plugin I cannot live without. It is built from the failures of spec-driven development, and it is oriented around human-in-the-loop decision making: scouts fetch the facts, the decisions come to me, and a task closes only when a command proves it. It has cut my token consumption, the model gets more of the work right, and it has changed the way I work with AI. I have no numbers for that yet. Evaluating this release is what I am doing next.
It is also the biggest of the four and the marketplace's first plugin. One unit of work gets one
.mightymodels/<slug>/ directory and one ticket.yml carrying its scope and model routing. The
work then moves through fixed stages: investigate, write down what is known, open a ticket, plan,
sprint, land the PR, review. Each stage has one job, writes one artifact, and reads ticket.yml
before it does anything else.
When the work merges, /prune-ticket compresses the directory into a 30-line archive and deletes
it. The unit of deletion is the unit of work, so working state never outlives the ticket that
produced it. None of it is ever tracked by git, by design: spec-driven development leaves slop
artifacts all over a repository, and they confuse the agents that read them later.
People ask me about the names, which are odd on purpose: ask-an-adult, merge-vader,
stick-the-landing, crashout. They are supposed to be memorable. /crashout is for when the
agent has earned a rant. It writes the rant down verbatim and checks whether the same failure has
happened before.
One of my favorite features is baton-pass. With AI it is hard to walk away from a session
without a penalty. baton-pass writes down the facts that live nowhere else and prints the
opening prompt for the next session, so the next one picks up where I stopped.
The mightymodels README has the loop diagram, every skill and worker, the hooks and where the state lives.
Skills for building the parts other plugins are made of: skills, subagents, hooks, rule files,
MCP servers, loops and whole plugins. There are eleven of them, and a vibe-code CLI that the
skills call to validate what they just wrote. It ships no agents.
It was called ai-engineer before 0.3.0. An install under the old name is not upgraded in place,
so remove it and install vibe-code.
The vibe-code README covers the skills, the CLI, the launcher and Windows.
Eight skills, four agents, a language server, two hooks, a CLI and an MCP server for Python
projects. The skills are build-web-service, enterprise-design, design-persistence and
design-tests (a service track that hands work along in that order), data-engineering and
data-analysis (a data track whose two skills hand work to each other), automate-task (a
script track) and tactical-slop-review, which every track hands to once code exists. Claude
offers a skill when the request matches its description, and you can run any of them as
/python-harness:<name>. tactical-slop-review reviews a PR, a branch or a codebase, writes
PYTHON-REVIEW.md or REFACTOR-PLAN.md, and never edits code. The four agents are pylens and
pep-talk, which are read-only and run on Haiku, reviewer, which cannot edit files, and
test-engineer, which writes the tests it is assigned. The ty language server gives Claude Code
type diagnostics for Python files and downloads ty through uv on its first start.
Two hooks come with it. One briefs each session on the project's Python toolchain. The other adds
a note when a Bash command runs plain python where uv run python was meant. Neither blocks a
command. The MCP server has eight tools: two read standard-library documentation from
docs.python.org, and six read the project Claude Code was started in.
Warning
As of 0.2.0 the skills, the agents, the language server, the hooks and the server have not yet been run inside a live Claude Code session. The python-harness README lists what is unconfirmed.
Four agents, five skills, three hooks, a language server and an MCP server for Terraform and IaC
repositories. The agents are terraform-advisor, which gives a second opinion on a Terraform
call, terraform-author, which writes Terraform inside a scope it is handed,
terraform-edit-reviewer, which checks a just-edited module for drift, and
terraform-researcher, which answers one question about a repository. The skills are
check-module, explain-plan, look-up-schema, review-module and trace-change-impact.
The hooks format a .tf or .tfvars file after Claude writes it, report terraform validate
and tflint findings after Claude reads one, and report the terraform, tflint and terraform-ls
versions when a session starts. The MCP server describes modules, runs fmt, validate and tflint,
digests plans, traces a change's impact and looks up provider schemas.
The plugin is developed in cc-terraform-harness,
not under plugins/, and the marketplace installs its v0.1.0 tag. Its
README lists every agent, skill and
hook.
plugins/ one directory per plugin; each carries its own manifest and skills
mightymodels/ the dev loop: nineteen skills, twelve worker agents
vibe-code/ authoring skills and the `vibe-code` CLI: eleven skills, no agents,
with its own tests and eval cases
python-harness/ Python engineering: eight skills, four agents, a ty language server, hooks,
the `python-harness` CLI and an MCP server, with its own tests
docs/ human documentation for the harness and the mightymodels plugin
scripts/ the quality gate (`quality.sh`), checkout setup, and a log helper
.claude-plugin/ the rygentic-harness marketplace manifest
| Page | What it covers |
|---|---|
| plugins/mightymodels/README.md | The mightymodels plugin: requirements, the loop, skills, workers, hooks, state |
| plugins/vibe-code/README.md | The vibe-code plugin: install, skills, CLI, launcher |
| plugins/python-harness/README.md | The python-harness plugin: install, skills, agents, language server, hooks, server, launcher, limitations |
| cc-terraform-harness README | The terraform-harness plugin: agents, skills, hooks, MCP server |
| docs/workflow.md | The full mightymodels loop, stage by stage, with diagrams |
| docs/skills.md | Every mightymodels skill: what it does and when it fires |
| docs/agents.md | The twelve workers and how models get routed |
| docs/state.md | The .mightymodels/ directory and ticket.yml |
| docs/claude-code.md | Running under Claude Code |
| CHANGELOG.md | What changed in each plugin, version by version |
A new plugin is a directory under plugins/ with a .claude-plugin/plugin.json naming it, plus
its skills/ and agents/. Add a matching entry to the marketplace manifest and the rest is
automatic: scripts/quality.sh runs claude plugin validate --strict for every directory under
plugins/ and for the marketplace root, which checks the manifests and the skill and agent
frontmatter.
A plugin developed in another repository gets a url source instead, the way terraform-harness
does: the repository's https:// clone URL, pinned to a release tag and that tag's commit. Use
url rather than a github source, because a github source is cloned over SSH and fails with
Permission denied (publickey) for anyone without a GitHub SSH key. claude plugin validate
checks that entry's shape and does not fetch the plugin, so the gate here never runs against its
files.
make check is the gate a change has to pass. It syncs the locked environment, runs pre-commit,
and runs scripts/quality.sh. Read CONTRIBUTING.md before you open a PR.
Evals are separate from the gate. Some mightymodels skills keep dated eval results in
plugins/mightymodels/skills/<skill>/evals/. The vibe-code eval cases live in
plugins/vibe-code/evals and run with claude plugin eval, which costs money.
For anything sensitive, see SECURITY.md.