English | 简体中文 | 繁體中文 | 日本語 | 한국어 | Русский | Français | Deutsch | Español | Português
stdagent is a lightweight, pure Go CLI that keeps a single .stdai/ directory as the source of truth for your project's AI configuration, then fans it out to 23 AI CLI tools with their native file formats, frontmatter dialects, and quirks handled for you.
Stop maintaining CLAUDE.md, AGENTS.md, GEMINI.md, .cursor/rules/, .windsurf/rules/, .clinerules/, .github/copilot-instructions.md, ... by hand. Edit once, sync everywhere.
- Single source — write
rules/skills/commands/references/subagentsonce in YAML frontmatter + Markdown. - Twenty-three targets — Claude Code, Codex, Cursor, GitHub Copilot, Windsurf/Devin, Gemini CLI, Aider, Cline, OpenCode, Roo Code, Crush, Amp, Warp, Factory, Continue.dev, Antigravity, Qwen Code, Pi, Kilo Code, Augment Code, Jules, Grok Build, Kimi Code.
- Spec-accurate — every output path, frontmatter dialect, and size limit is verified against the tools' official docs (last full audit: 2026-07); native Agent Skills directories everywhere they exist.
- Zero lock-in — the writer only touches a tiny whitelist of paths; backups before every sync;
cleanreverses everything. - Drift detection —
statusshows files modified outside stdagent;fixreapplies the source. - MCP — single
.stdai/standards/mcp.jsonfans out to.mcp.json/.cursor/mcp.json/.vscode/mcp.json - Monorepo aware — config lookup walks up from
cwd; works from any subdirectory. - Self-upgrading —
stdagent upgradepulls signed releases from GitHub with sha256 verification and atomic replace.
| Target | Primary outputs |
|---|---|
| Claude Code (Anthropic) | CLAUDE.md + .claude/{rules,skills,commands,agents}/ + .mcp.json |
| Codex (OpenAI) | AGENTS.md + .agents/skills/ + .codex/agents/*.toml (native subagents) |
| Cursor | .cursor/{rules/*.mdc,skills,commands,agents}/ + .cursor/mcp.json |
| GitHub Copilot | .github/{copilot-instructions,instructions,prompts,agents,skills}/ + .vscode/mcp.json |
| Windsurf / Devin (Cognition) | .windsurf/{rules,skills,workflows}/ + .devin/rules/ mirror |
| Gemini CLI (Google) | GEMINI.md + .gemini/skills/ + .gemini/commands/*.toml |
| Aider | reuses AGENTS.md (noop) |
| Cline | .clinerules/ (100/500/900 numeric prefixes) |
| OpenCode | .opencode/{skills,commands}/ |
| Roo Code | .roo/{rules,skills,commands}/ |
| Crush (Charmbracelet) | CRUSH.md + .crush/skills/ + crush.json skills registration |
| Amp (Sourcegraph) | AGENTS.md (inline) + .agents/skills/ |
| Warp | AGENTS.md (inline + nested) + .agents/skills/ |
| Factory (Factory.ai) | .factory/{rules,skills,commands,droids}/ |
| Target | Primary outputs |
|---|---|
| Continue.dev | .continue/{rules,skills,prompts}/ + nested rules.md |
| Antigravity (Google) | .agents/{rules,skills,workflows}/ |
| Qwen Code (Alibaba) | QWEN.md + .qwen/{rules,skills,commands}/ |
| Pi | .pi/skills/ + .pi/prompts/ |
| Kilo Code (kilo.ai) | .kilo/{rules,skills,command}/ + kilo.jsonc instructions registration |
| Augment Code | .augment/{rules,skills}/ |
| Jules (Google) | AGENTS.md |
| Grok Build (xAI) | AGENTS.md + .grok/skills/ |
| Kimi Code (Moonshot AI) | AGENTS.md + .agents/skills/ |
Each integration is documented under docs/targets/.
# Install (macOS / Linux)
curl -fsSL https://raw.githubusercontent.com/StringKe/std-agent/main/install.sh | sh
# Install (Windows PowerShell)
irm https://raw.githubusercontent.com/StringKe/std-agent/main/install.ps1 | iex
# Initialize in your project
cd your-project
stdagent init
# Edit .stdai/standards/rules/example.md, then sync to all enabled targets
stdagent sync
# Inspect / fix drift
stdagent status
stdagent fixProject already littered with CLAUDE.md / AGENTS.md / .cursor/rules/ / .clinerules/ / .github/copilot-instructions.md? Paste the prompt below into Claude Code / Codex / Cursor / Gemini CLI and it will reorganize everything into .stdai/standards/ for you.
Help me migrate this project from scattered AI configuration to std-agent. Please do:
1. Use Glob / Read to scan every existing AI rule file:
- Root: CLAUDE.md AGENTS.md GEMINI.md .cursorrules .windsurfrules .clinerules
- Subdirs: .claude/rules/ .claude/skills/ .claude/commands/ .claude/agents/
.cursor/rules/ .windsurf/rules/ .clinerules/ .continue/rules/
.github/copilot-instructions.md .github/instructions/
.rulesync/rules/ .codex/AGENTS.md
- In-repo nested CLAUDE.md (find . -name CLAUDE.md -not -path './.stdai/*')
2. Report an inventory: X rules / Y skills / Z commands / N nested CLAUDE.md,
and flag which files contain "project overview" content.
3. Propose a split plan, wait for my approval, then write files:
- Project overview (definition / stack / iron rules / maintenance flow)
-> .stdai/standards/root.md
- Each focused rule -> .stdai/standards/rules/<kebab-name>.md
- Skill package -> .stdai/standards/skills/<name>/SKILL.md (with scripts/ references/ subdirs)
- Slash commands -> .stdai/standards/commands/<name>.md
- Nested CLAUDE.md -> .stdai/standards/nested/<relative-path>/root.md
- Every file gets a frontmatter: type / name / description / priority / applyTo
4. No "refactoring" of original content. Keep every executable command, API endpoint,
error string, file path, line number. Allowed "optimizations": drop filler words,
merge duplicates, split oversized files, rename outdated tool names.
5. When done, tell me to run `stdagent sync` and remove legacy artifacts
(.rulesync/, .cursorrules single-file, etc.). DO NOT delete the files stdagent
itself produces (CLAUDE.md / AGENTS.md / .claude/rules/).
Full spec (frontmatter field table, root.md template, nested layout, rulesync mapping)
is in the `stdagent intro` command output.
Pipe straight into an LLM CLI as well:
stdagent intro | pbcopy # macOS: paste into AI chat
stdagent intro --json # for agent / automation integrations| Command | Purpose |
|---|---|
stdagent init |
Scaffold .stdai/ + config.toml + .stdaiignore + sample standards |
stdagent pull |
Update git-backed sources cached in .stdai/cache/ |
stdagent sync |
Core: pull → parse → convert → fan out |
stdagent fix |
Re-sync to repair drift (alias of sync) |
stdagent status |
Per-target drift + last sync time |
stdagent clean |
Remove generated files (preserves .stdai/) |
stdagent budget |
LLM context budget check (chars + token estimate) |
stdagent which <path> |
List rules / references applicable to a file (on-demand context routing for AI) |
stdagent explain |
Print std-agent 5 type semantics (rules/skills/commands/references/subagents) for AI |
stdagent intro |
Print a migration prompt for an LLM to convert your existing config |
stdagent upgrade |
Self-upgrade from GitHub Releases (sha256 + atomic replace) |
stdagent version |
Build info |
Every command supports --help. Full reference: docs/commands.md.
v0.0.4 introduced a three-layer transformer architecture: each target's Plan() delegates to one of 6 protocols (AgentsMD / ClaudeMD / Cursor / Clinerules / WindsurfStyle / Copilot), parametrized by a protocol.Adapter struct literal. Adding a new tool now costs ~25-35 lines instead of 145 (60-70% code dedup).
Graceful degradation: when a target doesn't natively support a std-agent type (e.g. references everywhere, subagents in amp / windsurf), stdagent falls back to subdirectory-isolated paths (<FallbackDir>/references/<name>.md) with frontmatter std-agent-type: <type> + HTML comment explainer, no std-agent-private prefixes.
A complete schema lives in docs/spec.md Part 1. The minimal shape:
---
type: rules # rules | skills | commands | references
name: coding-style
description: General coding style
priority: high # high | normal | low
targets: [claude-code, codex] # opt-in (or use exclude_targets to opt-out)
applyTo: ["**/*.go"]
alwaysApply: false
---
# Coding Style
Always use meaningful variable names...MCP servers (.stdai/standards/mcp.json):
{
"version": "1.0",
"servers": {
"github": { "type": "stdio", "command": "gh", "args": ["api"] },
"linear": { "type": "http", "url": "https://mcp.linear.app/sse" }
}
}.stdai/config.toml:
version = "1.0"
inject = true # inject "Generated by stdagent" footer in outputs
inject_whatis = true # add a one-line origin note inside skills
auto_pull = true # pull git sources on every sync
backup = true
backup_keep = 5
[targets]
claude-code = { enabled = true, convert = true }
codex = { enabled = true, convert = true }
cursor = { enabled = false, convert = true }
copilot = { enabled = false, convert = true }
windsurf = { enabled = false, convert = true }
gemini = { enabled = false, convert = true }
aider = { enabled = false, convert = true }
cline = { enabled = false, convert = true }
opencode = { enabled = false, convert = true }
continue-dev = { enabled = false, convert = true }
antigravity = { enabled = false, convert = true }
[sources.default]
url = "https://github.com/your-org/ai-standards.git"
branch = "main"
enabled = true
paths = ["standards/"]Full reference: docs/config-spec.md.
your-project/
├── .stdai/ Internal management area (single source of truth)
│ ├── config.toml One config file
│ ├── standards/ Authoring root
│ │ ├── rules/
│ │ ├── skills/
│ │ ├── commands/
│ │ ├── references/
│ │ └── mcp.json MCP servers (optional)
│ ├── cache/ Git source cache
│ ├── backups/ Auto-snapshot before each sync
│ └── state.json Runtime state
├── .stdaiignore gitignore-style globs to exclude source files
├── CLAUDE.md Fan-out: Claude Code
├── AGENTS.md Fan-out: Codex / Cursor fallback / Copilot agent / OpenCode / Antigravity
├── GEMINI.md Fan-out: Gemini CLI
├── .mcp.json MCP for Claude
└── .claude/ .codex/ .cursor/ .github/ .windsurf/ .gemini/ .clinerules/ .opencode/ .continue/ .agents/
Details: docs/file-structure.md.
When --config is omitted, stdagent walks up from cwd to find the nearest .stdai/config.toml. Run it from any subdirectory and it will locate the monorepo root automatically.
# Toolchain (mise + go + golangci-lint + gofumpt + git-cliff)
mise install
# Common tasks
mise run fmt # gofumpt + goimports
mise run lint # golangci-lint
mise run test # go test -race -cover
mise run check # fmt + lint + test in one go
mise run build # produces bin/stdagent
mise run run # go run ./cmd/stdagent- docs/spec.md — full spec: std-agent standard + 23-tool divergence + conversion strategy
- docs/prd.md — product requirements
- docs/architecture.md — module layout + data flow
- docs/commands.md — CLI command reference
- docs/conversion-rules.md — conversion matrix + frontmatter mapping
- docs/format-spec.md — frontmatter schema details
- docs/file-structure.md — directory conventions
- docs/roadmap.md — roadmap
- docs/targets/ — per-tool research notes
MIT — see LICENSE.