pipeflow-mcp
PiPeFlow passively tracks AI agent workflow states using a finite state machine, providing tools for declaring, advancing, and auditing pipelines with annotations and reports.
README
pipeflow-mcp
PipeFlow MCP Server —— AI Agent 任务的被动状态跟踪容器。
注意:本项目是实验 / demo 性质的学习作品,作者发布它用于代码评审与学习,真实场景下的实用价值尚未确证,不保证有用。
这是什么
pipeflow-mcp 是一个基于 Model Context Protocol (stdio JSON-RPC) 的 MCP 服务器。它把 agent 的工作流执行过程记录到一个由 FSM 驱动的 pipeline 账本里,提供:
- FSM 状态机:5 状态 × 5 action,
pending / running / succeeded / failed / skipped,由transitions库实现 - 审计完整性:
reset保留所有 annotations + reports,delete是 soft-delete,events表 append-only - 重试与时间预算:retry 计数器 lifetime,stage-level
time_budget_seconds派生time_status - 16 个
pipeflow_*工具:声明 / 状态推进 / 读取 / 注解 / 报告 / 仪表盘 / 健康检查,其中pipeflow_stage_advance是唯一的状态变更入口 - Cursor 分页:4 个 list 工具共享
base64-urlsafe(last_id + filter_fingerprint)不透明 cursor,filter 改变即 cursor 失效 - 静态 HTML 仪表盘:导出产物含 7 个 section —— Pipeline Overview / Decision Chain / Reflection Trace / Iteration Compare / Metric Analysis / Deliverables / Events Stream,Jinja2 autoescape 防 XSS,导出路径白名单
- 错误脱敏:客户端消息只含
pipeline_id/stage_id/ 错误类名,SQL 片段 / 文件路径 / 堆栈仅走 stderr 结构化日志 - Schema 校验:5 个 ReportType ——
decision / reflection / metric_report / deliverable / iteration_summary,对应 JSON Schema,启动时注册 - 类型化响应:8 个 Pydantic v2 view model,FastMCP
outputSchema自动填充字段契约,extra='forbid'锁定公开 JSON 契约
完整 API 文档见 docs/api.md。
状态
| 指标 | 值 |
|---|---|
| 版本 | v0.1.0 |
| MCP 工具 | 16 个 pipeflow_*,2 declare · 1 state_advance · 4 read · 3 annotation · 2 report · 3 dashboard · 1 health |
| 测试 | 584 passing,unit + integration + e2e + stdio smoke |
| 静态检查 | ruff ✅ · mypy --strict ✅ |
| Python | 3.11 - 3.14,运行时最低 3.11,CI 矩阵 3.13 + 3.14 |
| 存储 | SQLite WAL,单写者假设,多进程并发写不在设计范围内 |
安装与运行
MCP 客户端
从 GitHub 直接安装:
uvx --from git+https://github.com/AstralVoidZ/pipeflow-mcp pipeflow-mcp
或者从本地源码 checkout 启动:
git clone https://github.com/AstralVoidZ/pipeflow-mcp.git
cd pipeflow-mcp
uv venv --python 3.14 .venv
uv pip install -e ".[dev]"
uvx --from . pipeflow-mcp # dev 模式
Server 通过 stdio 监听 JSON-RPC,stdout 保留给 JSON-RPC,所有日志走 stderr。
客户端配置示例
Claude Code,~/.claude/mcp_servers.json:
{
"mcpServers": {
"pipeflow": {
"command": "uvx",
"args": ["--from", "git+https://github.com/AstralVoidZ/pipeflow-mcp", "pipeflow-mcp"],
"env": {
"PIPEFLOW_DB_PATH": "~/.pipeflow_mcp/state.db",
"PIPEFLOW_LOG_LEVEL": "INFO"
}
}
}
}
opencode,项目级 .opencode/config.json:
{
"mcp": {
"pipeflow": {
"type": "local",
"command": ["uvx", "--from", "git+https://github.com/AstralVoidZ/pipeflow-mcp", "pipeflow-mcp"],
"environment": {
"PIPEFLOW_DB_PATH": "~/.pipeflow_mcp/state.db",
"PIPEFLOW_LOG_LEVEL": "INFO"
},
"enabled": true
}
}
}
工具矩阵
| Category | Tool | 说明 |
|---|---|---|
| Declare | pipeflow_create |
一次性创建 pipeline + stages,支持 yaml_spec,与 name / stages / metadata 互斥 |
pipeflow_stage_append |
追加 stage 到已存在 pipeline | |
| State | pipeflow_stage_advance |
唯一状态变更入口,start / complete / fail / skip / retry |
| Read | pipeflow_get_state |
单 pipeline 当前状态 |
pipeflow_get_snapshot |
完整快照,含 reports 聚合 + time_status | |
pipeflow_list |
pipelines 列表,metadata_filter + cursor | |
pipeflow_history |
events 流,stage_id / event_types / since / until + cursor | |
| Annotation | pipeflow_annotation_upsert |
upsert 注解,决策记录 |
pipeflow_annotation_list |
注解索引,不含 data,3 层梯度 | |
pipeflow_annotation_get |
注解完整 payload | |
| Report | pipeflow_report_submit |
提交 5 种类型报告,schema 校验 |
pipeflow_report_list |
报告列表,type / iteration / cursor | |
| Dashboard | pipeflow_reset |
append-only reset,保留注解 + 报告 |
pipeflow_delete |
soft-delete,list 排除,history 保留 | |
pipeflow_report_export |
7 段 HTML 仪表盘导出,白名单路径 | |
| Health | pipeflow_health |
liveness + readiness + uptime,PRAGMA user_version DB 检查 |
完整参数 / 返回 schema / 错误清单见 docs/api.md。
配置
pipeflow-mcp不接受任何 CLI 参数,无--db-path之类的 flag。所有配置通过环境变量传入,客户端在 MCP server 配置的env块里设置即可。误传--db-path只会在 cwd 留下一个同名空文件。
| 变量 | 默认值 | 说明 |
|---|---|---|
PIPEFLOW_DB_PATH |
~/.pipeflow_mcp/state.db |
SQLite 数据库文件路径,首次启动自动 init_db |
PIPEFLOW_EXPORT_DIR |
~/.pipeflow_mcp/exports/ |
pipeflow_report_export 写入根,白名单校验 |
PIPEFLOW_LOG_LEVEL |
INFO |
日志级别,DEBUG / INFO / WARNING / ERROR |
PIPEFLOW_LOG_FORMAT |
text |
stderr 日志格式,设为 json 切换为结构化 JSON lines,字段 timestamp / level / logger / message / request_id / pipeline_id / stage_id / tool |
所有日志写入 stderr,stdout 保留给 JSON-RPC 通信。
架构
5 层严格单向依赖:
protocol (MCP tools / server.py)
→ view layer (Pydantic v2 view models, extra='forbid')
→ persistence orchestration (orchestration/ — advance, create, append, reset, delete, snapshot, annotate, reports, queries, ...)
→ { FSM layer (fsm/machine.py) , Spec layer (yaml_parser, report_registry, result_validator) }
→ data contract layer (models/*.py — dataclasses only)
→ error layer (errors/*.py)
侧模块,storage/ (DB + DAO + migrations) · pagination/ (cursor codec) · export/ (Jinja2 + 8 模板) · tools/ (16 薄封装) · config.py (env + logging)。
不变量
- stdout = JSON-RPC,stderr = logs,任何 stdout 写入会破坏 MCP 客户端
- DAO 是 SQL 边界,orchestration / fsm / spec / tools / pagination / export 永不 import
pipeflow_mcp.storage.dao - 单写者假设,SQLite WAL 在单 host + 单 writer 下安全,多进程并发写超出设计范围
user_version降级拒绝,永不静默 forward-migrate,启动时抛RuntimeError是正确行为- Annotation 列表不含
data,list 返回AnnotationSummary,data只能通过get取得 resetappend-only,保留 annotations + reports + 历史 events,只追加type='reset'event- View 单向投影,
tools → views → models是唯一允许方向,反向 import 被test_views_have_no_dao_imports强制禁止
开发
本地验证
ruff check src/pipeflow_mcp/ tests/
mypy src/pipeflow_mcp/
pytest tests/ -v --cov=pipeflow_mcp --cov-fail-under=85
打包
本地构建 wheel + sdist 到 release/ 目录,.gitignore 排除。
./scripts/build.sh # clean + lint + type + test + build
./scripts/build.sh --wheel # 只构建 wheel
./scripts/build.sh --check # 跳过 clean,直接基于当前 venv 构建
产物 release/pipeflow_mcp-<version>-*.whl 和 release/pipeflow_mcp-<version>.tar.gz,用于个人 GitHub Release 上传。
许可证
MIT,见 LICENSE。
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.