mcp-server-wtf
A flight recorder for Claude Code that captures tool calls and observations into a SQLite database, then classifies them into a distilled timeline for incident troubleshooting.
README
WTF (Why That Failed)
A flight recorder for incident troubleshooting inside Claude Code. WTF captures every tool call and manual observation into a durable SQLite database that survives context compaction, then enriches raw entries through a background classifier into a distilled timeline.
Prerequisites
- Claude Code CLI (
claude) jq(JSON processor)curlorwget- AWS credentials for Bedrock (optional — enables background classifier)
Installation
curl -fsSL https://raw.githubusercontent.com/Wave-Engineering/mcp-server-wtf/main/scripts/install-remote.sh | bash
This downloads a pre-compiled binary for your platform, installs the
PostToolUse hook, and registers the MCP server. No clone or runtime required.
Skills (/wtf, /wtf now, /wtf happened, /wtf imout) are delivered by
claudecode-workflow.
Verify the installation:
curl -fsSL https://raw.githubusercontent.com/Wave-Engineering/mcp-server-wtf/main/scripts/install-remote.sh | bash -s -- --check
Install a specific version:
curl -fsSL https://raw.githubusercontent.com/Wave-Engineering/mcp-server-wtf/main/scripts/install-remote.sh | bash -s -- --version v1.0.0
Development Installation
If you're working on the WTF server itself, clone the repo and use the local installer (requires Bun):
git clone https://github.com/Wave-Engineering/mcp-server-wtf.git
cd mcp-server-wtf
./scripts/install.sh
Quick Start
# Start troubleshooting
/wtf
# Investigate the issue... Claude records tool calls automatically
# Get the distilled timeline
wtf_happened
# Add your own observations
/wtf now "DNS resolver returning stale records"
Architecture Overview
WTF is a three-layer system:
/wtf, /wtf now Skills (user-facing entry points)
|
v
wtf_now, wtf_happened MCP Server (journal storage, retrieval,
wtf_freshell background classification)
^
|
PostToolUse hook Auto-capture (every tool call -> JSONL queue)
Layer 1 -- Skills: /wtf starts a troubleshooting session and activates
flight recorder mode. /wtf now adds manual journal entries.
Layer 2 -- MCP Server: A Bun + TypeScript server exposing four tools over stdio transport. Manages a SQLite database, ingests the hook queue, and runs a background classifier (Claude Haiku via AWS Bedrock) that categorizes entries as actions, breadcrumbs, theories, or noise.
Layer 3 -- PostToolUse Hook: A shell script that fires on every Claude Code
tool call, extracts relevant fields, truncates large values, and appends a JSON
line to .wtf/hook-queue.jsonl for ingestion.
Usage
Starting a Session
/wtf
Archives any prior incident and creates a fresh one. Prompts for an optional title, then puts Claude into flight recorder mode where significant observations are journaled automatically.
Recording Observations
/wtf now the health endpoint is returning 503s
/wtf now "theory: connection pool exhaustion under load"
/wtf now checked nginx logs — 502s started at 14:32
Adds a manual entry to the journal. Classification is handled by the background classifier.
Getting the Timeline
Call the wtf_happened MCP tool (Claude can invoke it directly, or you can ask
for it):
wtf_happened # summary (max 50 lines)
wtf_happened { detail: "full" } # all entries
Returns a distilled Markdown timeline:
## WTF Summary -- DNS Resolution Failure
**Duration:** 45 min | **Entries:** 127 raw, 34 distilled | **Status:** active
1. [12:03] BREADCRUMB -- Health endpoint returning 503s intermittently
2. [12:05] THEORY -- Connection pool exhaustion, idle timeout set to 0
3. [12:08] ACTION -- Set DB_POOL_IDLE_TIMEOUT=30s in .env
Suspending Recording
/wtf imout
Suspends the flight recorder without losing data. Useful when switching to non-troubleshooting work. The hook still fires but entries are filtered.
Generating a Runbook
wtf_happened also writes a runbook skeleton to .wtf/runbook.md that can be
refined into a reusable playbook.
Clearing / Starting Fresh
wtf_freshell
Archives the current incident and starts a new one. Previous entries are preserved in the database.
Configuration
Data Directory
All runtime data lives in .wtf/ relative to the project root (gitignored):
.wtf/
wtf.db SQLite database (WAL mode)
hook-queue.jsonl Hook queue (consumed by MCP server)
runbook.md Generated runbook skeleton
Tunable Values
| Parameter | Default | Location |
|---|---|---|
| Queue poll interval | 2,000 ms | queue.ts |
| Classifier poll interval | 5,000 ms | classifier/worker.ts |
| Classifier rate limit | 2,000 ms | classifier/worker.ts |
| Hook truncation limit | 4,096 bytes | scripts/hooks/wtf-post-tool-use.sh |
| Summary line cap | 50 lines | tools/happened.ts |
MCP Server Registration
The remote installer registers the compiled binary:
claude mcp add --scope user --transport stdio wtf-server -- ~/.local/bin/wtf-server
The development installer registers the Bun source directly:
claude mcp add --scope user --transport stdio wtf-server -- bun /path/to/mcp-server-wtf/index.ts
Hook Configuration
The PostToolUse hook is configured in ~/.claude/settings.json:
{
"hooks": {
"PostToolUse": [
{
"type": "command",
"command": "/absolute/path/to/scripts/hooks/wtf-post-tool-use.sh"
}
]
}
}
Uninstall
If installed via the remote installer:
curl -fsSL https://raw.githubusercontent.com/Wave-Engineering/mcp-server-wtf/main/scripts/install-remote.sh | bash -s -- --uninstall
If installed from a local clone:
./scripts/install.sh --uninstall
Both remove the MCP server registration and hook configuration. The
per-project .wtf/ data directories are preserved (they contain incident
history). To remove them:
rm -rf .wtf/
License
MIT -- see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。