claude-agents-core

A pack of general-purpose, reusable subagents for Claude Code.

Each agent is a focused, single-purpose Claude instance with its own context window, a tight tool allowlist, and a model chosen to match the job. They're the building blocks you reach for in almost any workflow — delegate the noisy or risky phase to a subagent and keep your main context clean.

These are intentionally generic: no project-specific tools, no external service dependencies. Drop the plugin in and the agents work anywhere.

Not included: code-reviewer / security-reviewer — those are well covered by the official code-review and security-review tooling. This pack fills the gaps around them.

Agents

Agent	Model	Tools	What it does
`adversarial-verifier`	sonnet	read-only	Independently tries to refute a claim/fix/finding before you trust it. Returns a grounded verdict.
`retriever`	haiku	read-only	Fast search across local files; returns the conclusion and excerpts, not file dumps.
`summarizer`	haiku	read-only	Distills logs, test output, and large files into a short, faithful summary.
`fact-checker`	sonnet	read-only + web	Verifies claims one at a time against primary sources, with citations.
`doc-writer`	sonnet	read + write	Drafts and revises docs grounded in the actual code.

Read-only agents (everything except doc-writer) cannot modify your files — by design. A reviewer that can edit tends to "fix" instead of report.

Install

Add the marketplace, then install:

claude plugin marketplace add fingerskier/claude-plugins
claude plugin install claude-agents-core@fingerskier-plugins

Usage

Once installed, the agents are available to the orchestrator and are routed to automatically when a request matches an agent's description. You can also invoke one explicitly, e.g.:

"Use the adversarial-verifier to check whether the null-pointer fix in parser.c actually covers the empty-input case."

"Have the retriever find every place we construct a database connection."

How it works

Each agent lives in agents/<name>.md — YAML frontmatter declares the agent (name, routing description, model, tool allowlist) and the markdown body is its system prompt. See the subagents docs.

Contributing

Issues and PRs welcome. New agents should be general-purpose (no project- or vendor-specific coupling), declare the minimum tools they need, and pick a model that fits the task (search → haiku, synthesis/review → sonnet).

claude-agents-core

Popularity

What's Inside

README