11 specialized AI agents + 8 productivity tools — all for Claude Code
한국어 README · Setup Guide · Cheatsheet · Harness Guide · MCP Guide · Integration · Contributing
A drop-in enhancement for Claude Code that gives you:
- 11 specialized sub-agents — each laser-focused on one job (design, code, review, test, security, harness design, pipeline orchestration…)
snippet— personal prompt manager; your best prompts one command awayclaude-handoff— save full session context and resume it in the next conversationclaude-cost— estimate and track Claude API spend before you run a promptclaude-review-diff— generate a structured code review prompt from your git diffclaude-remind— surface incomplete TODO items at session startclaude-harness— validate and generate AI harness definitions for specialist agentsclaude-pipeline— track multi-stage pipeline execution with quality gates and run reportsclaude-lessons— record why something failed and how it was fixed, searchable across sessions
All tools ship as Python CLIs and as Claude Code slash commands. The core tools also ship as a single compiled Rust binary (claude-tools).
Bilingual: All agents respond in the user's language — English and Korean (한국어) both fully supported.
You Orchestrator
│ │
└──► "Add OAuth login" ──► ├──► planner (architecture)
├──► database-expert (schema)
├──► implementer (code)
├──► reviewer (code review)
└──► tester (tests)
Benchmark scenario: fix a bug in a single TypeScript file. Dev rate: $60/hr · Pricing: Sonnet 4.6 $3/$15 · Opus 4.7 $5/$25 per 1M tokens.
| Metric | Vanilla Claude CLI | This Harness | Change |
|---|---|---|---|
| PR diff size | ~180 lines | ~25 lines | −86% |
| PR review time | ~18 min | ~4 min | −78% |
| First-pass approval rate | ~45% | ~85% | +89% |
| Rework rate | ~40% | ~10% | −75% |
| Bug detection (reviewer stage) | ✗ none | ✓ 42–48% caught | +∞ |
| API cost per task | ~$0.039 | ~$0.049 | +26% |
| Total engineering cost (API + dev time) | ~$0.32 | ~$0.10 | 3.1× cheaper |
API cost goes up slightly — but review time and rework drop so much that total cost (API + human time) is 3× lower per task.
Vanilla Claude CLI ████████████████████████████████ $0.32
├─ API ██ $0.04
├─ Review time ██████████████ $0.18
└─ Rework ████████ $0.10
This Harness ████████ $0.10
├─ API ███ $0.05
├─ Review time ████ $0.04
└─ Rework █ $0.01
Vanilla ██████████████████████████████████████ 180 lines
Harness ████ 25 lines (surgical edit protocol)
Vanilla ████████████████████ 18 min
Harness ████ 4 min
Data sources: Google DORA Report 2025 · Anthropic Engineering Blog · Stripe PR velocity data · DevToolsAcademy AI code review benchmark 2025
npm install -g @anthropic-ai/claude-code
claude # authenticate on first runOption A — One-line installer (recommended, macOS / Linux)
curl -fsSL https://raw.githubusercontent.com/BcKmini/claude-code-use/main/install.sh | bashOption B — Specific version
curl -fsSL https://raw.githubusercontent.com/BcKmini/claude-code-use/main/install.sh | bash -s -- --version v1.0.0Option C — Clone and build from source
# Windows
git clone https://github.com/BcKmini/claude-code-use.git
cd claude-code-use
powershell -ExecutionPolicy Bypass -File setup-agents.ps1# macOS / Linux
git clone https://github.com/BcKmini/claude-code-use.git
cd claude-code-use
make install # agents + slash commands + Python tools
make install-rust # optional: Rust binary (requires cargo)claude
/agents # → 9 agents listed
/snippet list # → built-in snippets
| # | Agent | Model | Autonomy | Job |
|---|---|---|---|---|
| 00 | orchestrator | Opus | L2 | Breaks down requests and delegates to sub-agents |
| 01 | planner | Opus | L1 | Architecture & design decisions — read-only |
| 02 | implementer | Sonnet | L2 | Writes and edits code |
| 03 | reviewer | Sonnet | L1 | Bug, security, quality, performance review — read-only |
| 04 | tester | Sonnet | L2 | Unit, integration, E2E test authoring |
| 05 | security-auditor | Opus | L1 | OWASP Top 10 audit — read-only |
| 06 | performance-optimizer | Sonnet | L2 | Bottleneck analysis and optimization |
| 07 | database-expert | Sonnet | L2 | Schema design, queries, migrations |
| 08 | documenter | Haiku | L3 | README, API docs, inline comments |
| 09 | harness-designer | Opus | L1 | Designs tight/loose/adaptive AI harnesses for automation |
| 10 | pipeline-orchestrator | Opus | L2 | Manages multi-stage pipelines with context isolation |
All agents are bilingual — they detect the user's language and respond in English or Korean (한국어).
Each agent carries only the context relevant to its role. Parallel execution (planner + security-auditor simultaneously) cuts wall-clock time.
Autonomy (L0 = human does everything → L4 = fully autonomous) is a separate axis from harness type — it says how much human checking a task needs, not how constrained the output is. See Autonomy Levels.
See AGENT-CHEATSHEET.md for 24+ ready-to-use prompts.
Eight productivity tools that fill the gaps Claude Code doesn't cover out of the box.
| Command | What it does |
|---|---|
/snippet |
Run, save, list personal prompt templates |
/handoff |
Save/load session context across conversations |
/cost |
Estimate and track API spend |
/review-diff |
Code review prompt from current git diff |
/remind |
Surface pending TODO items at session start |
/harness |
Design and validate AI harness definitions |
/pipeline |
Run and track multi-stage AI pipelines |
/lessons |
Record and recall why past tasks failed and how they were fixed |
Save your best Claude prompts by name and recall them in one command.
snippet list # all snippets
snippet run full-pipeline | claude # pipe to Claude
snippet save myfix "Fix {{BUG}} in {{FILE}}" # template variables
snippet search security
snippet export my-backup.json/snippet list
/snippet run full-pipeline
/snippet search security
20 built-in snippets including full-pipeline, code-review, security-audit, write-tests, refactor, db-schema, and more. See snippets/defaults.json.
Save the full context of a session and load it back in the next one.
claude-handoff save --note "OAuth done, next: email verification"
claude-handoff load | claude # resume immediately
claude-handoff list
claude-handoff clean --days 30/handoff save
/handoff load
/handoff list
What a handoff captures: git branch, last 5 commits, working-tree status, diff stat, TODO.md contents, your note, and a ready-to-paste Resume Prompt.
Know what a prompt will cost before you run it. Track actual spend from session logs.
claude-cost estimate --snippet full-pipeline --agents 9
claude-cost month
claude-cost set-budget 20.00/cost estimate full-pipeline
/cost month
/cost agents
| Model | Input (per 1M) | Output (per 1M) |
|---|---|---|
| Opus | $15.00 | $75.00 |
| Sonnet | $3.00 | $15.00 |
| Haiku | $0.25 | $1.25 |
Generate a structured code review prompt from your current git changes and pipe it straight into Claude.
claude-review-diff # review unstaged changes
claude-review-diff --staged # review staged changes
claude-review-diff --base main # compare branch to main
claude-review-diff --focus security # security-only review
claude-review-diff | claude # pipe directly to Claude/review-diff
/review-diff --staged
/review-diff --base main --focus security
Focus options: security · performance · correctness · style · tests · all
Output groups findings by severity: Critical → Major → Minor → Nit
Scan TODO.md / TASKS.md / CLAUDE.md for incomplete checkboxes and print a session-start prompt.
claude-remind # show all pending tasks + resume prompt
claude-remind --quiet # count only
claude-remind | claude # pipe to Claude to resume/remind
/remind --quiet
Typical end-of-session / start-of-session workflow:
# End of session
claude-handoff save --note "Auth done, next: email verification"
# Start of next session
claude-remind | claude # see pending tasks
claude-handoff load | claude # full context restoreValidate agent harness definitions and generate harness templates from the command line.
claude-harness check-all # validate all agents in agents/
claude-harness validate agents/09-harness-designer.md # validate one agent
claude-harness template tight my-specialist # print a tight harness template
claude-harness template adaptive my-orchestrator # print an adaptive harness template/harness design automate slow query detection and patching
/harness validate agents/03-reviewer.md
/harness types
Checks performed on each agent:
- Role is clearly scoped
- Output format is constrained
- Forbidden actions are listed
- Tools list is minimal
- Bilingual language support present
Track multi-stage AI pipeline execution, log stage results, and generate Markdown run reports.
claude-pipeline init slow-query-fix # create and activate pipeline
claude-pipeline stage "detection" start
claude-pipeline stage "detection" pass --note "found 3 slow queries"
claude-pipeline stage "patch-gen" start
claude-pipeline stage "patch-gen" warn --note "1 query had no safe fix"
claude-pipeline status # show live status
claude-pipeline report # markdown run report
claude-pipeline list # all saved pipelines/pipeline run analyze slow queries and generate patches with review loop
/pipeline status
/pipeline stages
Record why something failed and how it was fixed, so the next session (or agent) doesn't repeat it. Unlike claude-handoff (session-scoped, prunable), lessons accumulate indefinitely and are searchable by tag or keyword.
claude-lessons add --title "Migration timed out" --tags db,migration \
--symptom "ALTER TABLE locked prod for 4min" \
--cause "no lock_timeout set" \
--fix "added SET lock_timeout='2s' before DDL"
claude-lessons list --tag db
claude-lessons search lock_timeout
claude-lessons context | claude # pipe recent lessons into a new session/lessons add
/lessons list --tag db
/lessons context
Typical workflow:
# During/after debugging a tricky failure
claude-lessons add
# Start of a new session touching the same area
claude-lessons context --tag db | claudeAll tools compiled into one zero-dependency binary — no Python required.
cd rust && cargo build --release
# or: make install-rust
claude-tools snippet list
claude-tools handoff save --note "done"
claude-tools cost month
claude-tools watch # live cost monitor (real-time)
claude-tools watch --interval 5
claude-tools env # environment health checkclaude-tools env output:
Claude Code Environment
✓ ANTHROPIC_API_KEY sk-ant-…abcd
✓ ~/.claude/ exists
✓ ~/.claude/agents/ 9 agents installed
✓ ~/.claude/commands/ 5 commands: snippet, handoff, cost, review-diff, remind
✓ handoffs 3 saved, latest: 20250608-143022.md
✓ sessions 4 projects, 12 session files
make help # list all targets
make install # agents + slash commands + Python tools
make install-rust # build and install Rust binary
make build # cargo build --release
make test # fast smoke tests (no extra deps)
make tox # full Python test matrix (py38-py313) + lint + fmt-check
make msrv # verify the Rust crate builds on its declared MSRV
make lint # clippy (Rust)
make fmt # rustfmt + ruff format (scoped — see CONTRIBUTING.md)
make status # git log + tool install check
make env # Claude environment health check
make clean # remove build artifactsClaudecode-Agent/
├── Makefile ← build / install / test / clean
├── tox.ini ← Python test matrix + lint + fmt-check
├── pyproject.toml ← ruff config (scoped to this repo's style)
├── setup-agents.ps1 ← Windows quick installer
├── setup-agents.sh ← macOS / Linux quick installer
│
├── agents/ ← agent definitions → ~/.claude/agents/
│ ├── 00-orchestrator.md
│ ├── 01-planner.md
│ ├── 02-implementer.md
│ ├── 03-reviewer.md
│ ├── 04-tester.md
│ ├── 05-security-auditor.md
│ ├── 06-performance-optimizer.md
│ ├── 07-database-expert.md
│ ├── 08-documenter.md
│ ├── 09-harness-designer.md ← NEW harness architect
│ └── 10-pipeline-orchestrator.md ← NEW pipeline manager
│
├── .claude/commands/ ← slash commands → ~/.claude/commands/
│ ├── snippet.md
│ ├── handoff.md
│ ├── cost.md
│ ├── review-diff.md
│ ├── remind.md
│ ├── harness.md ← NEW /harness
│ ├── pipeline.md ← NEW /pipeline
│ └── lessons.md ← NEW /lessons
│
├── snippets/
│ └── defaults.json ← 20 built-in prompt templates
│
├── tools/
│ ├── snippet.py ← prompt manager CLI
│ ├── claude-handoff.py ← session continuity CLI
│ ├── claude-cost.py ← cost estimator CLI
│ ├── claude-review-diff.py
│ ├── claude-remind.py
│ ├── claude-harness.py ← NEW harness validator + template gen
│ ├── claude-pipeline.py ← NEW pipeline tracker + reporter
│ ├── claude-lessons.py ← NEW failure/lessons-learned log
│ ├── install-tools.ps1 ← Windows tool installer
│ └── install-tools.sh ← macOS/Linux tool installer
│
├── tests/ ← pytest suite for tools/*.py
│ ├── conftest.py ← run_tool / home / git_repo fixtures
│ └── test_*.py ← one file per tool
│
├── rust/claude-tools/src/
│ ├── main.rs
│ ├── snippet.rs
│ ├── handoff.rs
│ ├── cost.rs
│ ├── watch.rs ← live cost monitor
│ ├── env.rs ← NEW environment health check
│ └── colors.rs
│
├── examples/
│ └── mcp-lessons-server.py ← NEW MCP server example
│
└── docs/
├── SETUP.md / SETUP.ko.md
├── AGENT-CHEATSHEET.md / .ko.md
├── HARNESS-GUIDE.md / .ko.md ← NEW harness design guide
├── MCP-GUIDE.md / .ko.md ← NEW MCP server guide
├── INTEGRATION.md / .ko.md
├── CONTRIBUTING.md / .ko.md
└── CLAUDE.md / .ko.md
| Situation | Command |
|---|---|
| After finishing a step | /compact |
| Switching tasks completely | /clear |
| Check current spend | /cost |
| Watch live cost | claude-tools watch |
| Check environment | claude-tools env |
| Resume with pending tasks | claude-remind | claude |
| Full session restore | claude-handoff load | claude |
| Recall past failures for this area | claude-lessons context --tag X | claude |
Agents not showing in /agents
ls ~/.claude/agents/*.md # files must be present
# Restart Claude Code after installingAgent Teams not working
# Windows — should print 1
[System.Environment]::GetEnvironmentVariable("CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS","User")# macOS / Linux — should print 1
echo $CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMSTool not found after make install-tools
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc && source ~/.bashrcmake not available on Windows
winget install GnuWin32.Make- New tool idea? → Feature Request
- Bug? → Bug Report
- New snippet? → add to
snippets/defaults.jsonand send a PR
See CONTRIBUTING.md for the full guide.