Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/florianbruniaux/claude-code-plugins/eval-agentsnpx skills add FlorianBruniaux/claude-code-plugins --skill eval-agentsgit clone --depth 1 https://github.com/FlorianBruniaux/claude-code-pluginsWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00093 | $0.03497 |
| Opus 5 | $0.00046 | $0.01749 |
| Sonnet 5 | $0.00019 | $0.00699 |
| Haiku 4.5 | $0.00009 | $0.00350 |
Grade A, and why
eval-agents scanned grade A with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Unrestricted tool accesslowExcessive agency
A wildcard tool grant or "run any command" leaves no least-privilege boundary at all.
- **`tools: "*"` wildcard**: full session access, treat as equivalent to no tools field unless the system prompt explicitly justifies why unrestricted tools are needed (e.g., a general-purpose agent) Downgraded: this mod is about security review, or the phrase is quoted, so it is likely naming the pattern rather than instructing it.
How it starts
The opening of the file, as written. The whole thing — 305 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Agent Evaluator
Discover all agents in scope, score each one across five criteria, then run an interactive session to confirm or improve them one by one.
Agents are not just scripts: they are callable units selected by orchestrators based on their description. A vague description silently breaks multi-agent workflows. The goal here is not just scoring; it is leaving every agent correctly scoped, correctly modeled, and safe to call from an orchestrator.
When to Use
- First time auditing an agent fleet before wiring it into an orchestration pipeline
- An orchestrator keeps selecting the wrong agent for a task
- After copying agents from another project or importing a plugin
- A new agent was added; checking whether it conflicts with existing ones
- Periodic hygiene: "do all these agents still do something distinct?"
Key Concepts
Agent file locations
| Location | Scope | Committed? |
|---|---|---|
.claude/agents/<name>.md |
Project (flat file) | Yes |
.claude/agents/<name>/AGENT.md |
Project (directory-based) | Yes |
~/.claude/agents/<name>.md |
User-level | No |
Plugin agents/*.md |
Per plugin | Yes (in plugin) |
Both flat files and directory-based agents are valid. The agent name is always the filename or directory name, not the name: field in frontmatter (though they should match).
Frontmatter fields
| Field | Required | Notes |
|---|---|---|
name |
No | Display label. The runtime key is always the filename/directory. |
description |
Yes (for dispatch) | Used by orchestrators to select agents. Vague = wrong agent selected. |
model |
No | Defaults to session model if absent. Explicit is always better. |
tools / allowed-tools |
No | Both spellings parse correctly. Defaults to session tools if absent (risky). |
disallowed-tools |
No | Blocks specific tools even if granted by the session. |
effort |
No | low, medium, high, xhigh, max. Overrides session effort. |
context |
No | fork runs the agent in isolated subagent context. |
argument-hint |
No | Hint shown during autocomplete. |
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 305 lines · 93 tokens per session scan A bace8cc4004a
eval-agents is a skill published in the GitHub repository FlorianBruniaux/claude-code-plugins (40 stars, last pushed 3mo ago), licensed MIT. It adds 93 tokens to every session and 3,497 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it A with 1 finding (unrestricted tool access). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
orchestrate
Decompose a multi-file task, fan out to parallel workers in dependency-ordered waves, and reconcile their results — with a one-writer-per-file guarantee. The third orchestration mode (parallel fan-out) alongside sequential and iterative-retrieval. Rides the platform Workflow tool as the engine; the skill is the lead …
extract
Turn a repeatable workflow observed this session into a PROPOSED reusable skill — draft it, evaluate it against the rubric, and stage it for your approval. Never writes into skills/. Opt-in; not wired into the default pipeline. Use when a Stop candidate appears or a workflow felt worth keeping.
operate
Supervise a long-running, unattended agent run — an autonomous loop or a scheduled wake-up — that self-checkpoints against the repo's eval skills so it can't silently drift. Rides the platform's /loop + /schedule; adds halting guardrails (iteration cap, wall-clock budget, drift) and durable run state. Opt-in; not…
self-iteration
系统自迭代能力，通过数据驱动的模式识别生成改进提案.
experience-creator
Author runnable OpenExpertise experience files — heterogeneous executable graphs (agent / skill / tool / dataset / experience nodes) backed by a persistent blackboard. Use this skill whenever the user wants to create, write, build, scaffold, or design an OpenExpertise experience: "make an experience for X", "create an…
scaffolding
Generates boilerplate following project conventions. Use when asked to create a new component, module, service, or scaffold code.