flowmind

🧠 FlowMind

The AI Agent That Learns How You Work

Memory and workflow automation for MCP, Codex, and Claude Code.

中文 | Quick Start | Demo | Skills | Integration | Security | Changelog

What It Is

FlowMind is a memory layer for repeatable developer workflows.

Use it when you already have tools and want a better execution surface:

• Route recurring tasks through reusable skills
• Capture explicit feedback and reuse it later
• Run the same workflow from CLI, Codex, or MCP clients
• Audit and change microservice governance across Alibaba Cloud, Huawei Cloud, and Baidu Cloud through guarded workflows

Best-fit use cases today:

• Log and trace investigation
• Code review with fixed standards
• Data validation and internal tool orchestration

Strong Features

• Reusable skill routing turns repeated work into one-command workflows instead of prompt rewrites
• Guarded workflow execution adds dry-run review, approval tokens, and resumable follow-up chains
• Multi-cloud service governance supports audit, diagnosis, and policy changes with provider-aware rollback previews
• Local workflow history preserves execution state so interrupted chains can be resumed instead of restarted
• MCP exposure keeps FlowMind usable from CLI, Codex, and other integrated clients

Quick Start

npm install -g flowmind
flowmind doctor
flowmind demo log-audit
flowmind demo code-review
flowmind skills
flowmind process --skill log-audit "query the latest error logs"

Guarded follow-up workflows can be reviewed, executed, listed, and resumed from the CLI:

flowmind workflow --chain follow-up-chain.json
flowmind workflow --chain follow-up-chain.json --execute --approval-token <token>
flowmind workflow-history
flowmind workflow-history --status failed --resumable
flowmind workflow-resume --execution-id <id> --approval-token <token>

If you want Codex-friendly JSON output:

flowmind-codex skills
flowmind-codex --skill log-audit "查询 traceId abc123 的日志"

If you prefer npx:

npx flowmind@latest doctor
npx flowmind@latest demo log-audit
npx flowmind@latest skills

Runtime supports Node.js 18 and newer. Contributor tooling runs best on Node.js 20.19 or newer.

Configuration-Free Demo

Run the built-in mock demo before connecting any provider or MCP server:

flowmind demo log-audit
flowmind demo code-review
flowmind demo workflow
flowmind demo log-audit --json

This shows the shape of a log-audit result without requiring SLS, databases, internal platforms, or AI provider credentials.

Capability Scope

Area	Works out of the box	Requires configuration
Skill discovery	flowmind skills	None
Mock demo	flowmind demo log-audit, flowmind demo code-review, flowmind demo workflow	None
Learning inspection	flowmind learn --assets	Local FlowMind state
Offline design analysis	flowmind analyze "summarize local code and design history" --offline	Local material and repository paths
Log, data, docs, workflow skills	Skill routing and result contracts	MCP servers, provider credentials, or adapters
Multi-cloud governance	Guarded workflow contracts	Provider-specific MCP governance adapters

Local State

FlowMind stores local config and learning data under FLOWMIND_HOME/.flowmind. If FLOWMIND_HOME is not set, it falls back to your home directory.

To import local SDD-Agent resources and scene mappings, run:

flowmind resource --sync-sdd-agent

To inspect the currently learned skill feedback and local material packages:

flowmind learn --assets
flowmind learn --assets --skill log-audit --limit 20 --json

MCP clients can use flowmind_learning_assets for the same sanitized, read-only view.

Without an AI model, FlowMind can analyze synchronized learning material, Lake/Lakebook documents, and repository code locally:

flowmind analyze "summarize the historical design and code flow" --offline
flowmind analyze "explain the renewal rules" --source-dir /data/source --repo-dir /work/service --json

The result includes an evidence-backed flow, document/code summaries, confidence, limitations, and test focus. MCP clients can call flowmind_offline_analyze.

When an AI model is available, an ordinary “analyze this business flow” request remains in AI skill selection or AI answering. The deterministic offline skill is forced only by explicit offline/no-AI wording or --skill offline-design-analysis.

For safety, cross-workspace auto-sync is disabled by default. Initialization only auto-syncs when you explicitly set FLOWMIND_ENABLE_SDD_AGENT_AUTO_SYNC=1.

Files that may contain credentials or workflow state are written with private permissions on POSIX filesystems. See SECURITY.md and docs/security-model.md.

Why It Works

• Skills provide a stable execution path instead of ad hoc prompts
• Feedback becomes reusable state instead of chat history
• CLI and MCP entrypoints make the workflow scriptable

A Real Example

flowmind process --skill log-audit "查询 traceId abc123 的日志"
flowmind "下次用表格格式"

On the next similar request, FlowMind can reuse that explicit formatting preference from local learning data.

Explicit skill choices are learned as reusable routes as well:

flowmind process --skill offline-design-analysis "analyze the current contract limit validation flow"

Later exact or highly similar requests prefer the learned skill. In interactive chat, this should use offline-design-analysis binds the previous request. Analyze, summarize, or explain flow requests are treated as analysis; execute, deploy, or run pipeline actions remain auto-flow operations.

The TUI shows skill, MCP, Git, and configuration readiness, including a Bind: skill->component line for skills that still need resources.

Read By Goal

• Need the full product walkthrough: docs/guide.md
• Need the skill capability matrix: docs/skills-catalog.md
• Need to add a provider or adapter: docs/provider-adapter-guide.md
• Need the upgraded skill contracts and chaining patterns: docs/upgrade-contracts.md
• Need Claude Code, Codex, Cursor, or MCP setup: docs/integration-guide.md
• Want to see terminal output first: demo/DEMO.md
• Want release history: CHANGELOG.md

Why This Homepage Is Short

The npm package page works best as an entry page, not as the entire manual. Detailed usage modes, system explanation, and architecture notes now live in the linked docs.

Release Notes

Package release details are maintained outside the published README.

Contributing

PRs are welcome. Start with CONTRIBUTING.md.

Before opening a PR, run:

npm run quality
npm run audit:prod

License

MIT. See LICENSE.