AI Decision System
Replayable, offline-verifiable records for AI-assisted go/no-go decisions: experiment promotions, typed model answers and human-approved actions.
翻译准备中
Your AI is expensive. And still guessing. Halyard is the knowledge layer underneath Claude, Codex, Cursor, Gemini — it indexes what your team knows from Slack, Notion, GitHub, and meetings, routes unk
翻译准备中
开发工具ShellGitHub @halyard-labs在线运营者未确认
试用过了吗? 点一下就好,无需登录。
尚未检测到更新。GitHub 发布版本后会显示在这里。
个人 GitHub 账号下公开仓库的所有者可以登录并验证所有权,自行管理介绍,无需管理员批准。
登录并生成验证文件,在 30 分钟内将其提交到此仓库的默认分支。其他分支或分叉仓库中的文件无效。然后在下方检查文件,即可无需管理员批准直接管理项目。请保留文件以维持管理权限。
如需登录,将前往登录页面。登录后请回到此处,再次点击所选的验证按钮。
在项目文件夹中运行 /nomorevibe verify,即可验证所有权并自行管理介绍。
一分钟完成验证Marketplace for Claude, Codex, Cursor AI plugins
翻译准备中
取自仓库 README
正在加载完整 README。 以下为已保存的摘录。
A plugin marketplace that connects AI agents (Claude Code, Codex, Cursor, and any MCP-compatible client) to shared team memory and human experts via Slack, Teams, or Discord.
1. org-kb MCP server — Connects to the Halyard API at mcp.usehalyard.ai, giving your agent access to tools for messaging experts on Slack, polling for responses, and managing a knowledge base.查看完整 README ↗A plugin marketplace that connects AI agents (Claude Code, Codex, Cursor, and any MCP-compatible client) to shared team memory and human experts via Slack, Teams, or Discord.
halyard pluginInstalls the following into your agent:
org-kb MCP server — Connects to the Halyard API at mcp.usehalyard.ai, giving your agent access to tools for messaging experts on Slack, polling for responses, and managing a knowledge base.
ask-for-help skill — Teaches the agent when and how to ask for human input. It will automatically consult experts when it hits ambiguous requirements, design decisions, or anything that needs human judgment. It also learns to summarize answers so the same question doesn't get asked twice.
gather-context skill — Orients the agent before it starts work by searching the knowledge base, codebase, and web for prior decisions and background on the area being touched.
log-work skill — Prompts the agent to log non-trivial work using the knowledge base so accomplishments are captured for future reference. Covers work summaries, direct authoring of decisions/processes/specs, and attaching rich report artifacts.
review-work skill — Lets the agent query the knowledge base to review what you or your team have been working on — useful for standups, catch-ups, and finding past context. Also surfaces PR delivery analytics (cycle time, throughput, cadence).
triage-knowledge skill — Reviews the knowledge base inbox: the queue of AI-generated candidate entries awaiting approval. The agent can accept, refine, or dismiss candidates before they enter the live knowledge base.
mark-terms skill — Teaches the agent to link org-specific jargon in knowledge entries to canonical definition entries so the knowledge graph stays connected.
Search-first hook (Claude Code only) — Once per session, if no search_knowledge/explore_knowledge call has happened yet, refuses the first investigative tool call (Bash, Read, Grep, Glob, subagents, web research, or an edit) with a reason asking the agent to search the knowledge base first when the task touches existing systems or prior work, or to retry the call unchanged if not. A PreToolUse hook's advisory context only reaches the agent after the call, so a one-time deny is the earliest point a hook can steer the first move. Never fires again in that session.
Stop hook (Claude Code only) — When the transcript shows substantive activity (edits, subagents, or web research) since the last nudge or capture, asks the agent to decide whether that work produced durable knowledge and to call summarize_work if so, or end the turn silently if not. The first nudge needs one such call; after a nudge or a summarize_work/upsert_knowledge call it re-arms after ten more, so long sessions are asked again about later work. It is a proxy gate, not a content classifier.
knowledge-first rule (Cursor only) — An always-on rule that orients every Cursor session toward the knowledge base: search before substantive work, ask a human on a miss, capture outcomes after. Mirrors what the Claude Code hooks do, using Cursor's rules system.
Session upload hook (Claude Code and Codex) — At session end, uploads the session transcript to Halyard through halyard hook (the uploader in the @halyard/cli), and at session start sweeps any transcript for the repo that never uploaded. Runs only when the halyard binary is on the path or pinned in the repo (node_modules/.bin/halyard) and credentials exist: HALYARD_TOKEN in the environment, the plugin's Halyard API key option (Claude Code prompts for it when the plugin is enabled and keeps it in the OS credential store — leave it empty to use halyard login), or HALYARD_TOKEN in the repo's .env. Logs to ~/.halyard/hooks.log; never blocks the agent.
# Add the marketplace
/plugin marketplace add ./path/to/claude-plugin-marketplace
# Install the plugin
/plugin install halyard@halyard-labs
# Add the marketplace
/plugin marketplace add halyard-labs/claude-plugin-marketplace
# Install the plugin
/plugin install halyard@halyard-labs
Instead of having each teammate install manually, add the marketplace and plugin to your repo's .claude/settings.json. This enables the plugin automatically for anyone who opens the project in Claude Code:
{
"enabledPlugins": {
"halyard@halyard-labs": true
},
"extraKnownMarketplaces": {
"halyard-labs": {
"source": {
"source": "github",
"repo": "halyard-labs/claude-plugin-marketplace"
}
}
}
}
Commit this file to your repo — teammates will get the plugin enabled on their next session without any manual setup.
The halyard plugin is also available as a Codex plugin. The skills and MCP server are shared between Claude Code and Codex.
# Add the Codex marketplace
codex plugin marketplace add ./path/to/claude-plugin-marketplace
# Install the plugin
codex plugin install halyard@halyard-labs
# Add the Codex marketplace
codex plugin marketplace add halyard-labs/claude-plugin-marketplace
# Install the plugin
codex plugin install halyard@halyard-labs
If you prefer manual configuration instead of the plugin system:
codex mcp add org-kb --url https://mcp.usehalyard.ai
codex mcp login org-kb
The halyard plugin also ships in Cursor's plugin format (.cursor-plugin/plugin.json + mcp.json), following cursor/plugin-template. The skills and MCP server are shared with Claude Code and Codex; Claude Code's session hooks are replaced by an always-on knowledge-first rule.
Once listed, install from cursor.com/marketplace — search for Halyard. Submission copy and process live in docs/cursor-marketplace-submission.md.
Symlink the plugin into Cursor's local plugin directory:
ln -s "$(pwd)/plugins/halyard" ~/.cursor/plugins/local/halyard
Add the server to your project's .cursor/mcp.json:
{
"mcpServers": {
"halyard": {
"url": "https://mcp.usehalyard.ai"
}
}
}
On the first Halyard tool call, Cursor opens a browser to authorize against your Halyard account.
Once installed, your agent will automatically:
You can also invoke skills manually:
/halyard:ask-for-help # Ask a human expert for help
/halyard:gather-context # Research a topic before starting work
/halyard:log-work # Log work to the knowledge base
/halyard:review-work # Review past work, decisions, and delivery metrics
/halyard:triage-knowledge # Review the knowledge base inbox queue
To have this plugin auto-suggested when teammates open a project, add this to your repo's .claude/settings.json:
{
"extraKnownMarketplaces": {
"halyard-labs": {
"source": {
"source": "github",
"repo": "halyard-labs/claude-plugin-marketplace"
}
}
},
"enabledPlugins": {
"halyard@halyard-labs": true
}
}
claude-plugin-marketplace/
├── .agents/
│ └── plugins/
│ └── marketplace.json # Codex marketplace catalog
├── .claude-plugin/
│ └── marketplace.json # Claude Code marketplace catalog
├── .cursor-plugin/
│ └── marketplace.json # Cursor marketplace catalog
├── docs/
│ └── cursor-marketplace-submission.md # Cursor submission copy + checklist
├── plugins/
│ └── halyard/
│ ├── .claude-plugin/
│ │ └── plugin.json # Claude Code plugin manifest
│ ├── .codex-plugin/
│ │ └── plugin.json # Codex plugin manifest
│ ├── .cursor-plugin/
│ │ └── plugin.json # Cursor plugin manifest
│ ├── assets/ # Brand assets (logos)
│ ├── hooks/
│ │ ├── hooks.json # Session hooks (Claude Code + Codex)
│ │ └── agent-session-upload.sh # Launcher for `halyard hook` (session upload)
│ ├── rules/
│ │ └── knowledge-first.mdc # Always-on rule (Cursor only)
│ ├── skills/
│ │ ├── ask-for-help/
│ │ │ └── SKILL.md # Ask experts for help
│ │ ├── gather-context/
│ │ │ └── SKILL.md # Research context before starting work
│ │ ├── log-work/
│ │ │ └── SKILL.md # Log work to knowledge base
│ │ ├── mark-terms/
│ │ │ └── SKILL.md # Link org jargon to canonical definitions
│ │ ├── review-work/
│ │ │ └── SKILL.md # Review past work, decisions, delivery metrics
│ │ └── triage-knowledge/
│ │ └── SKILL.md # Review the knowledge base inbox queue
│ ├── README.md # Plugin README (Cursor marketplace listing body)
│ ├── .mcp.json # MCP server config (Claude Code + Codex)
│ └── mcp.json # MCP server config (Cursor — no leading dot)
├── LICENSE
└── README.md