Executable discipline for AI coding agents (Claude Code, Cursor, Copilot, Gemini CLI, OpenCode, Codex) - skills the agent loads on demand and applies while it works, not prose guides it ignores.
Agent Plugins is a plugin marketplace for Claude Code and OpenAI Codex.
Works with any compatible agent (Claude Code, Cursor, Copilot, Gemini CLI, OpenCode, Codex, and more). Each command is independent - run the one(s) you want:
# code-intelligence (this repo's inline skill)
npx skills add https://github.com/antonbabenko/agent-plugins
# terraform-skill (its own repo)
npx skills add https://github.com/antonbabenko/terraform-skill/plugin marketplace add antonbabenko/agent-plugins
/plugin install code-intelligence@antonbabenko
/plugin install terraform-skill@antonbabenko
/plugin install claude-delegator@antonbabenkocodex plugin marketplace add antonbabenko/agent-pluginsThen run codex, open /plugins, select Agent Plugins, and install
code-intelligence, terraform-skill, or claude-delegator.
For other hosts, expand below.
Gemini CLI
External plugin (terraform-skill) installs as an extension from its own repo:
gemini extensions install https://github.com/antonbabenko/terraform-skillThe inline code-intelligence has no standalone repo. Clone the Agent Plugins
repo and point Gemini at plugins/code-intelligence per Gemini's
skill-discovery docs, or install it through Claude Code or Codex.
Cursor
# external plugin (terraform-skill) - from its own repo
git clone https://github.com/antonbabenko/terraform-skill.git ~/.cursor/skills/terraform-skill
# inline plugin (code-intelligence) - from this repo
git clone https://github.com/antonbabenko/agent-plugins.git
ln -s "$(pwd)/agent-plugins/plugins/code-intelligence" ~/.cursor/skills/code-intelligenceCursor auto-discovers skills from .agents/skills/ and .cursor/skills/.
Copilot
git clone https://github.com/antonbabenko/terraform-skill.git ~/.copilot/skills/terraform-skill
git clone https://github.com/antonbabenko/agent-plugins.git
ln -s "$(pwd)/agent-plugins/plugins/code-intelligence" ~/.copilot/skills/code-intelligenceCopilot auto-discovers skills from .copilot/skills/.
OpenCode
git clone https://github.com/antonbabenko/terraform-skill.git ~/.agents/skills/terraform-skill
git clone https://github.com/antonbabenko/agent-plugins.git
ln -s "$(pwd)/agent-plugins/plugins/code-intelligence" ~/.agents/skills/code-intelligenceOpenCode auto-discovers skills from .agents/skills/, .opencode/skills/, and
.claude/skills/.
Codex (OpenAI) - clone fallback
Prefer the Codex marketplace block above. Plain skills-directory fallback:
git clone https://github.com/antonbabenko/terraform-skill.git ~/.agents/skills/terraform-skill
git clone https://github.com/antonbabenko/agent-plugins.git
ln -s "$(pwd)/agent-plugins/plugins/code-intelligence" ~/.agents/skills/code-intelligenceCodex auto-discovers skills from ~/.agents/skills/ and .agents/skills/.
Update with git pull in each clone.
Antigravity
git clone https://github.com/antonbabenko/terraform-skill.git ~/.antigravity/skills/terraform-skill
git clone https://github.com/antonbabenko/agent-plugins.git
ln -s "$(pwd)/agent-plugins/plugins/code-intelligence" ~/.antigravity/skills/code-intelligenceUpdate with git pull in each clone.
Manual (Claude Code - symlink a local clone)
git clone https://github.com/antonbabenko/agent-plugins.git
mkdir -p ~/.claude/plugins
# inline plugins live under plugins/<name>/
ln -s "$(pwd)/agent-plugins/plugins/code-intelligence" ~/.claude/plugins/code-intelligenceClaude Code autodiscovers skills/<name>/SKILL.md on next launch. Edits to the
clone are picked up live. External plugins (e.g. terraform-skill) are not in
this repo - install them from their own repo instead.
Stops blind text-replace renames and "the tool is broken" guesses: the agent picks language-server vs text vs fuzzy search correctly for the task, and says so on the first line when it has to swap one tool for another.
tldr - what changes with the plugin:
| Prompt | Without the plugin | With the plugin |
|---|---|---|
Rename var.tags to var.resource_tags across the module |
Blind grep/replace - misses position scoping, collides with same-named symbols in child modules, no post-edit validate |
terraform-ls has no rename provider, so it runs the safe manual workflow: findReferences at an anchored position -> fresh per-file Read -> edit -> validate |
Find every reference to aws_s3_bucket.this before changing it |
One regex; may over/under-match, then claims "found all" | Position-anchored findReferences; on an uninitialized workspace it says so on line 1 and discloses the rg fallback |
rg looks missing so the agent reaches for grep |
Silent tool swap - you never learn coverage dropped | Proves the tool is really absent first; states the substitution on the first line |
The semantic rows above need a language server installed (here terraform-ls).
Without one the plugin still helps: it discloses the rg fallback on the first
line instead of pretending the text search was exhaustive. Check your setup
with /code-intelligence:setup-code-intelligence.
/plugin install code-intelligence@antonbabenkoTry:
I need to change var.vpc_id - what references it?Where does local.name_prefix come from?What can I set on this aws_s3_bucket resource?Is output.cluster_endpoint used anywhere before I change its type?Did you really find all the references, or just text matches?
Check your setup (Claude Code): /code-intelligence:setup-code-intelligence
Routes a Terraform or OpenTofu request to its real failure mode - identity churn, secret exposure, blast radius, state corruption - before generating code, instead of emitting plausible-looking HCL that breaks on apply.
tldr - what changes with the plugin:
| Prompt | Without the plugin | With the plugin |
|---|---|---|
| Create a VPC module | Plausible HCL, no tests, no version guards | Routes to the failure mode first; module + native tests + version-aware guards |
| Configure S3 backend with state locking | May emit DynamoDB locking (outdated) | TF 1.10+ S3 native use_lockfile; flags DynamoDB as no longer required |
| Change a resource's address | Text rename -> destroy/recreate on apply | moved block + plan shows 0 destroy; blast radius checked before generating code |
/plugin install terraform-skill@antonbabenko
# or, on any Agent Skills host:
npx skills add https://github.com/antonbabenko/terraform-skillTry:
Create a VPC module with native testsConfigure S3 backend with state locking
Source and detail: github.com/antonbabenko/terraform-skill. The full per-host install list lives in that repo's README.
Gives the agent five GPT (Codex) and/or Gemini expert subagents - it delegates the hard call (architecture, plan review, scope, code review, security) instead of guessing alone, and synthesizes the result rather than pasting it raw.
tldr - what changes with the plugin:
| Prompt | Without the plugin | With the plugin |
|---|---|---|
| Is this auth flow secure? | One model's single take | Security Analyst expert reviews; verdict synthesized, not raw |
| Review this migration plan | Self-review, same blind spots | Plan Reviewer expert validates before you execute |
| Get GPT and Gemini to agree on this design | Manual back-and-forth | consensus iterates GPT + Gemini + Claude to a signed-off plan |
Each expert runs advisory (read-only) or implementation (workspace-write).
Maintained fork of jarrodwatts/claude-delegator (upstream inactive); MIT.
/plugin install claude-delegator@antonbabenko
/claude-delegator:setupRequires the Codex CLI and/or
Gemini CLI; /setup guides you through
it.
Bundled commands:
/claude-delegator:setup- configure Codex/Gemini MCP servers + rules/claude-delegator:uninstall- remove MCP config, rules, and aliases/claude-delegator:ask-gpt- one-shot GPT (Codex) second opinion/claude-delegator:ask-gemini- one-shot Gemini second opinion/claude-delegator:ask-both- GPT + Gemini in parallel, synthesized/claude-delegator:consensus- iterate GPT + Gemini + Claude to consensus
Use consensus when the plan must be right - GPT + Gemini + Claude iterate
until all three sign off, ideal for high-stakes planning and design. Use the
ask-* commands for a quicker single or parallel opinion when you just want a
fast second take.
/setup can also install short aliases (/ask-gpt, /ask-gemini,
/ask-both, /consensus); opt-in, never overwrites an existing command.
Source and detail: github.com/antonbabenko/claude-delegator.
- Honest by construction. Any tool substitution or skipped step is stated on the first line of the response, so you can trust what the agent reports.
- Token-lean. A short
SKILL.mdroutes to reference files that load only when the task needs them. The agent does not carry the whole guide in context. - Portable. One discipline across Claude Code, Cursor, Copilot, Gemini CLI, OpenCode, and Codex, with no per-host retraining.
- Composable and pinned. Generic skills give the base discipline; domain skills extend it. Each plugin is released independently, so upgrading one never moves another.
Built and maintained by Anton Babenko - LinkedIn, X/Twitter.
The model and the test requirements are in CONTRIBUTING.md.
Apache 2.0. See LICENSE.