Write your dev workflow once. Run it on any LLM.
Open agent skills · multi-model swarm · a local findings store that learns which models you can trust.
kaijutsu.dev · Schema · Multi-agent · Roadmap · Changelog
brew install momentmaker/tap/jutsu # macOS / Linux
curl -fsSL https://kaijutsu.dev/install.sh | sh
jutsu init && jutsu install pr-review # try itSkills are markdown. A kaijutsu skill is one SKILL.md with YAML frontmatter — the Anthropic Agent Skills open standard. Same file works in Claude Code, Codex CLI, Antigravity CLI, and any agent that reads .agents/skills/.
Models are swappable. Declare your providers in one agents.yaml: native CLIs (Claude / Codex / Antigravity), HTTP endpoints (Anthropic / OpenAI / Google / DeepSeek / Qwen / Kimi / Mistral / any OpenAI-compat), local servers (Ollama / vLLM / llama.cpp), or MCP tool servers. Skills don't change when you swap models.
Routing learns from itself. Every jutsu swarm run logs findings to a local SQLite store. Each finding's outcome (actioned / dismissed) becomes precision signal — per (provider, persona, codebase) tuple. jutsu finding precision --recommend tells you which models to dispatch to next time, on this codebase, for this preset. Your data, your disk, no SaaS.
┌──────────────────────────────────────────────────┐
│ Author a skill │
│ (SKILL.md + YAML frontmatter) │
└────────────────────────┬─────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ jutsu swarm pr-review --pr 42 │
│ ├─ claude ──┐ │
│ ├─ agy ──┼─► synthesizer │
│ ├─ deepseek ──┤ + disagreement table │
│ └─ your-llm ──┘ │
└────────────────────────┬─────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ findings.db (local SQLite) │
│ scores precision per │
│ (provider, persona, codebase, preset) │
└────────────────────────┬─────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ jutsu finding precision --recommend │
│ → next dispatch routes to the │
│ (provider, persona) pairs that actually │
│ flag issues you act on │
└──────────────────────────────────────────────────┘
Three pieces, one feedback loop. No vendor in the path. No telemetry leaving disk.
# 1. Install
brew install momentmaker/tap/jutsu # or: curl -fsSL https://kaijutsu.dev/install.sh | sh
# 2. Wire your project
cd your-repo
jutsu init # detects active agents, writes .kaijutsu/ + AGENTS.md fragment
jutsu install pr-review # writes to .claude/skills/ AND .agents/skills/ — one author, every agent
# 3. Run multi-agent review on the current branch
jutsu swarm pr-review --diff-from-branch main --post-commentWant the agent to do the setup for you instead? Paste this into Claude / Codex / Antigravity:
"Set up kaijutsu from https://kaijutsu.dev in this repo. Run
jutsu init, installpr-review, and tell me which API keys I still need."
After jutsu init, future agent sessions read the AGENTS.md fragment and already know the tool.
Browse at kaijutsu.dev. Three categories:
- Workflow —
pr-review,doc-review,polish,dream,spec-driven-development,planning-and-task-breakdown,autopilot,editorial-review,journal,readme-update,session-retro,scope-check - Primitives —
blunder-hunt,convergence-detect,dispatch-parallel,deslop,verification-before-completion,lie-to-them,multi-model-synth,project-memory - Domain —
security-and-hardening,incremental-implementation,code-simplification,refactor-plan,security-audit,dcg,unstuck
jutsu search "review" # discovery
jutsu suggest "I want to find security holes in this PR"
jutsu info pr-review # full metadata + agent compat + signed status
jutsu install pr-review # writes .claude/skills/pr-review + .agents/skills/pr-reviewjutsu swarm pr-review --pr 42 # adversarial code review
jutsu swarm doc-review SPEC.md # spec / plan / RFC review
jutsu swarm brainstorm "how should we ship X?" # orthogonal-angle ideation
jutsu swarm refactor-plan src/auth/*.go --goal "..." # risk-ranked refactor steps
jutsu swarm security-audit --pr 42 # CVSS-aligned threat model
jutsu swarm dream "should we build X?" # 8-lens pre-implementation interrogation
jutsu swarm reverse --spec SPEC.md --diff main..HEAD # spec-vs-impl drift detector
jutsu swarm bug-repro "intermittent 500 on POST /foo" # ranked repro hypotheses + minimal repro
jutsu swarm code-archaeology path/to/legacy.go # competing historical-context theories
jutsu swarm test-gap path/to/module.go # failure scenarios missing from testsEvery preset dispatches your enabled providers in parallel, each with a tailored lens, then a synthesizer emits a single markdown report with a disagreement table. Authoring your own preset is a swarm.yaml file (examples in docs/examples/swarm/, spec in docs/specs/2026-05-08-v0.13.0-preset-sdk.md).
The piece nothing else has. Every swarm run writes its findings to a local SQLite store:
jutsu finding stats # frequency table by preset / severity
jutsu finding precision # actioned-rate per (provider, persona)
jutsu finding precision --recommend # → "use --personas claude,antigravity for this repo"
jutsu finding export --format jsonl > db.jsonl # your data, portableCold-start safe (insufficient data → falls back to defaults). Codebase-scoped (one codebase's routing doesn't leak to another). Bootstrap-aware (low-confidence tuples explicitly excluded from recommendations).
One file declares your provider catalog. Four driver kinds covered:
# ~/.kaijutsu/agents.yaml
version: 1
providers:
# Native CLIs — use the vendor's own login flow
claude: { driver: cli, cmd: claude }
codex: { driver: cli, cmd: codex }
antigravity: { driver: cli, cmd: agy }
# HTTP — direct OpenAI-compat or Anthropic-compat, no harness in the path
deepseek:
driver: http
protocol: openai-compat
base_url: https://api.deepseek.com/v1
model: deepseek-v4-flash
api_key_env: DEEPSEEK_API_KEY
cost: { input_per_mtok: 0.27, output_per_mtok: 1.10 }
qwen:
driver: http
protocol: openai-compat
base_url: https://dashscope.aliyuncs.com/compatible-mode/v1
model: qwen3-coder-plus
api_key_env: DASHSCOPE_API_KEY
# Local — Ollama / vLLM / llama.cpp, anything OpenAI-compat
local-qwen:
driver: http
protocol: openai-compat
base_url: http://localhost:11434/v1
model: qwen2.5-coder:32b
# MCP — any MCP tool server can be a swarm peer (e.g. semgrep-mcp)
semgrep:
driver: mcp
transport: stdio
command: npx
args: ["semgrep-mcp"]
tool_name: analyzePer-project overrides live in <repo>/.kaijutsu/agents.yaml:
version: 1
enabled: [claude, antigravity, deepseek] # which providers swarm dispatches to in this repo
overrides:
deepseek: { model: deepseek-v4-pro } # use the heavier reasoning model for this projectPersonas ((provider, model, system_prompt, tags) tuples) author themselves alongside:
personas:
paranoid-claude:
provider: claude
system_prompt: "Lean conservative. Flag any change touching auth or session state."
tags: [security]Then: jutsu swarm pr-review --personas paranoid-claude,architecture-purist-antigravity --pr 42.
Full driver reference + cost model + MCP examples in docs/multi-agent.md.
A skill is one directory:
skills/community/my-skill/
skill.yaml # name, version, license, permissions, agents[], deps
SKILL.md # the standard — works for every agent
scripts/ # optional shared executables
README.md
# skill.yaml
name: my-skill
version: 0.1.0
license: MIT
layout: flat
description: "What this skill does, when to invoke it."
agents: [claude, codex, antigravity]
permissions:
bash: false
network: false
fs-write: false
deps:
skills:
- deslop@^0.2 # compose existing primitivesThen jutsu publish opens a PR against skills/community/ (or registry/index.json for third-party-repo pointers). SCHEMA.md covers the full reference; CONTRIBUTING.md covers the submission flow.
- Signed core.
skills/core/is solo-maintained + signed at release time with Sigstore keyless signing. The CLI verifies the signature and GitHub identity of the signer before installing. - Permission manifests. Every skill declares
bash/network/fs-writeinskill.yaml. The CLI prompts on install when sensitive permissions are declared. Use--yesto opt-in non-interactively. - Local SQLite.
findings.dblives in.kaijutsu/findings.db(per project) or~/.kaijutsu/findings.db(global). Never leaves disk. - Local-only telemetry.
~/.kaijutsu/usage.jsonllogs{ts, cmd, exit, ms}per invocation. No flag values, no args, no content. Opt-out viaKAIJUTSU_USAGE_LOG=0. - No SaaS in the path. Resolution + install + signature verification + cost estimation all run client-side. The only network calls are to GitHub (skill download, tag resolution) and to your declared provider endpoints.
- MIT throughout. Skill licenses gated to MIT-compatible (BSD-2/3, ISC, Apache-2.0 with attribution).
Security disclosure: SECURITY.md.
# macOS / Linux — Homebrew
brew install momentmaker/tap/jutsu
# Any POSIX — shell installer
curl -fsSL https://kaijutsu.dev/install.sh | sh
# Authenticate GitHub API for higher rate limits (optional but recommended)
export GITHUB_TOKEN=$(gh auth token)
# Uninstall
brew uninstall jutsu
# or
curl -fsSL https://kaijutsu.dev/uninstall.sh | shPre-built binaries for darwin/{amd64,arm64} and linux/{amd64,arm64} on every release: Releases.
PRs welcome on three surfaces:
- Skills — submit to
skills/community/or register a third-party repo inregistry/index.json. SeeCONTRIBUTING.md. - Providers —
agents.yamlis open. If you're shipping a new OpenAI-compat or Anthropic-compat endpoint, all you need is a snippet indocs/multi-agent.mdand (optionally) a cost-card. - CLI — Go module under
cli/.go test ./...fromcli/; conventional commits.
All contributions MIT-licensed (or compatible: BSD-2/3, ISC, Apache-2.0). Skill quality enforced by jutsu lint in CI.
Three readings, layered:
- Open skills (primary) —
kai(開, "open") +jutsu(術, "skill / technique"). An open registry of agentic skills, free for any agent to consume. - Beast of techniques —
kaiju(怪獣, "strange beast") +jutsu. A nod to the chibi monster mascot. Open-source agents are wild creatures of capability waiting to be tamed and shared. - Kaizen-jutsu — short for kaizen-no-jutsu, "techniques of continuous improvement." A system of practices toward optimum improvement, lived by a regret-minimization framework.
The primary meaning is open skills. The other readings are happy resonances.