Make your coding agent plan better, remember more, verify its own work, and stay safe — in under 5 minutes.
A thin, pluggable enhancement layer for the agents you already use.
One MCP server · a clean CLI · a local dashboard.
Quick Start · Features · How it works · Tools · Full Guide · 中文
Coding agents are powerful but unreliable in predictable ways: they go off-track, over-engineer, lose context between sessions, plan weakly, and occasionally run something dangerous. Most fixes mean adopting a heavy new platform.
AgentBoost takes the opposite bet. It's a thin layer that plugs into your existing agent and does five things well:
The agent keeps thinking. AgentBoost gives it structure, memory, guardrails, and feedback — all stored in a plain
.agentboost/folder you can read, diff, and edit.
- 🔌 No lock-in — works with Claude Code, Cursor, OpenCode, Codex, and any MCP client
- 🔑 No API keys, no cloud — your agent does the reasoning; data never leaves your machine
- ⏱️ Useful in 5 minutes — one command to init, one to connect, one to open the dashboard
| Feature | What it does | |
|---|---|---|
| 🧭 | Structured Planning | Turn a natural-language goal into a clean Spec + small, individually verifiable steps. Refine iteratively; track per-step status. |
| 🧠 | Context & Memory | Persistent project memory — decisions, architecture, preferences, and hard-won learnings. Keyword search, importance-aware pruning, one-call context injection. |
| ✅ | Automatic Verification | Auto-detects your test / typecheck / lint / build commands (npm, tsc, pytest, go, cargo, make), runs them, and returns clear pass/fail to drive fix loops. |
| 🛡️ | Safety & Permissions | Three permission levels + dangerous-op detection (rm -rf, git push --force, DROP TABLE, DELETE without WHERE…) with human-confirmation gates and an audit log. |
| 🧩 | Parallel Sub-agents | A bounded, concurrency-capped task runner with simple result aggregation for fan-out work. |
| 📊 | Local Web Dashboard | See the current phase, plan progress, memory, verification results, and a live activity log — edit memory and permissions right from the browser. |
agentboost gui opens a clean, localhost-only dashboard that shows everything at a glance:
┌────────────────────────────────────────────────────────────────┐
│ ⚡ AgentBoost my-project [ executing ] Permission ▾ ↻│
├────────────────────────────────┬───────────────────────────────┤
│ Current Plan │ Verification [Run checks]│
│ Ship search feature │ ✅ test npm run test │
│ ▓▓▓▓▓▓▓░░░ 3/5 done │ ✅ lint npm run lint │
│ [x] Add DB index │ ❌ build npm run build │
│ [~] Wire /search API ├───────────────────────────────┤
│ [ ] Add integration tests │ Project Memory [+ Add]│
├────────────────────────────────┤ [decision] Use SQLite │
│ Safety Check │ [architecture] REST, not gRPC │
│ > rm -rf build/ ⛔ critical │ [preference] 2-space indent │
└────────────────────────────────┴───────────────────────────────┘
💡 Real screenshots/GIFs belong in
docs/screenshots/— run the dashboard againstexamples/demo-projectto capture them.
# 1 · Initialize AgentBoost in your project (creates .agentboost/)
npx agentboost init
# 2 · Open the dashboard
npx agentboost gui
# 3 · Print the MCP config for your agent
npx agentboost adaptersConnect Claude Code — drop this into .mcp.json at your project root and restart:
{
"mcpServers": {
"agentboost": {
"command": "npx",
"args": ["-y", "agentboost", "mcp"],
"env": { "AGENTBOOST_ROOT": "." }
}
}
}Then just tell your agent: "Plan this with AgentBoost first." It calls plan_start → plan_create, records decisions with memory_add, runs verify_run after changes, and checks risky commands with safety_check — all visible live in the dashboard.
👉 Full walkthrough (install, every agent, examples, troubleshooting): docs/guide.md
| Agent | Integration | Config |
|---|---|---|
| Claude Code | MCP server | adapters/claude-code → .mcp.json |
| Cursor | MCP server | adapters/cursor → .cursor/mcp.json |
| OpenCode / Codex | MCP server | adapters/opencode |
| Any MCP client | stdio MCP | npx -y agentboost mcp |
| No agent (CLI) | terminal | every feature works standalone |
┌─────────────────────────────┐
your agent ────▶│ AgentBoost MCP server │
(Claude/…) │ plan · memory · verify · │
│ safety · status │
└──────────────┬──────────────┘
you ──▶ CLI ───────────────────┤ one shared core
you ──▶ GUI ───────────────────┘ │
▼
.agentboost/ (plain files)
config.yaml · memory.json · plans/ · state.json · audit.log
The MCP server, CLI, and dashboard are three faces of one core, reading and writing a single plain-file data directory. Nothing is hidden — inspect it, diff it in git, or hand-edit it.
The agent-facing surface is 13 focused tools:
| Category | Tools |
|---|---|
| Planning | plan_start · plan_create · plan_refine · plan_set_step_status · plan_get |
| Memory | memory_add · memory_search · memory_context · memory_edit |
| Verification | verify_run |
| Safety | safety_check · safety_set_level |
| Status | status |
agentboost/
├── src/
│ ├── core/ planning · memory · verification · safety · audit · sub-agents · facade
│ ├── mcp/ MCP server (primary agent integration)
│ ├── gui/ local dashboard: http server + static SPA (no build step)
│ └── cli/ command-line tool
├── adapters/ Claude Code · Cursor · OpenCode config templates
├── examples/ runnable demo project
├── docs/ bilingual operation guide + screenshots
└── test/ node --test suite
Zero-install via npx agentboost <command>, or install globally:
npm install -g agentboostRequirements: Node.js ≥ 18. Minimal dependencies — MCP SDK, commander, yaml, zod.
npm install
npm run build # compile TS + copy static assets to dist/
npm test # run the test suite (node --test)
npm run dev -- status # run the CLI from source via tsxContributions welcome — see CONTRIBUTING.md.
- Real dashboard screenshots & a short demo GIF
-
plan_starttemplates for common task types (bugfix, feature, refactor) - Richer verification (coverage thresholds, custom gates)
- Optional standing-context sync with
AGENTS.md/CLAUDE.md - Publish to npm
MIT © AgentBoost contributors
AgentBoost 是一个轻量、可插拔的增强层,让你现有的编码 Agent(Claude Code、Cursor、OpenCode、Codex)规划更清晰、记忆更持久、能自动验证、操作更安全。
它不是又一个庞大的 Agent 平台,而是一层薄薄的增强:Agent 负责思考,AgentBoost 负责提供结构(计划)、持久化(记忆)、护栏(安全)和反馈(验证) —— 全部存放在项目里可读、可 diff、可手改的 .agentboost/ 纯文本目录。
- 🔌 无锁定:兼容 Claude Code、Cursor、OpenCode、Codex 及任何 MCP 客户端
- 🔑 无需 API key、无需联网:推理交给你的 Agent,数据不出本机
- ⏱️ 五分钟见效:一条命令初始化,一条接入,一条打开面板
| 功能 | 说明 | |
|---|---|---|
| 🧭 | 结构化规划 | 把自然语言目标拆成清晰的 Spec + 可逐条验证的小步骤,支持迭代细化与进度跟踪 |
| 🧠 | 上下文与记忆 | 持久化项目记忆(决策/架构/偏好/经验),关键词搜索、按重要度智能修剪、一键上下文注入 |
| ✅ | 自动验证 | 自动识别并运行测试 / 类型检查 / lint / 构建,给出清晰通过/失败,驱动修复循环 |
| 🛡️ | 安全与权限 | 只读 / 有限写入 / 完全访问 三级权限 + 危险操作检测(rm -rf、强推、DROP TABLE 等)+ 人工确认 + 审计日志 |
| 🧩 | 并行子 Agent | 有并发上限的任务执行器 + 结果聚合 |
| 📊 | 本地 Web 面板 | 实时查看阶段、计划进度、记忆、验证结果与操作日志,可在浏览器里直接编辑 |
npx agentboost init # 在项目里初始化(生成 .agentboost/)
npx agentboost gui # 打开本地面板
npx agentboost adapters # 打印你所用 Agent 的 MCP 接入配置详细中文操作指南(中英双语)见 docs/guide.md。