state-machine
Enables validating, simulating, and analyzing finite state machines via MCP tools for step execution, path finding, and schema inspection.
README
Step Switch
Step Switch turns a finite business workflow into executable data. Given one Machine Spec, a snapshot, and an event, it deterministically accepts or rejects the transition, returns the next snapshot, and emits symbolic effect intents. It never executes those effects.
The repository includes one shared TypeScript core, a CLI, six read-only MCP tools in a Codex plugin, and a browser editor/simulator.
It intentionally stays smaller than a complete statechart runtime: guards are explicit caller-supplied facts, effects are symbolic intents, Agent inputs are inline and closed-world, and no surface evaluates user code.
Install the Codex plugin
The public source-distribution channel is the GitHub repository marketplace.
After the v0.1.0 release is available:
codex plugin marketplace add tetracoralla/state-machine --ref v0.1.0
codex plugin add state-machine@state-machine
Restart ChatGPT or Codex, open a new task, and use a concrete Machine Spec. For
example, ask it to validate examples/order.machine.yaml, test whether an event
is legal from a snapshot, or simulate an event sequence. An ordinary supported
request should require one machine.* tool call.
The repository includes .agents/plugins/marketplace.json and the committed
prebuilt server, so plugin users do not need npm, TypeScript, or a build step.
See the current OpenAI plugin packaging documentation
for marketplace and host behavior.
Develop from source
npm ci
npm run check
Node.js 22 or later and npm 10 or later are required. The Node package is intentionally private and is not an npm distribution channel.
Run the editor
npm ci
npm run build
npm run start:ui
Open http://127.0.0.1:4317. The workspace ships with an order lifecycle
example. Edit the YAML definition, inspect validation and topology, choose
explicit guard outcomes, run events, inspect the trace and effect intents, find
a route to a target state, and import or export the spec.
For development with live reload:
npm run dev:ui
Use the CLI
npm run build
node dist/node/adapters/cli.js validate examples/order.machine.yaml --pretty
node dist/node/adapters/cli.js step examples/order.machine.yaml \
--event '{"type":"PAYMENT_SUCCESS","payload":{"amount":128,"payment_id":"pay_1024"}}' \
--guards '{"payment_amount_matches":true}' \
--pretty
node dist/node/adapters/cli.js simulate \
examples/order.machine.yaml \
examples/order.events.yaml \
--pretty
node dist/node/adapters/cli.js path \
examples/order.machine.yaml \
completed \
--pretty
The CLI also supports inspect and diff. Run it without arguments for the
complete command summary. Exit code 0 means the operation completed, including
ordinary transition rejections and unreachable paths. Exit code 1 means the
machine or operation result is invalid; exit code 2 is reserved for command,
file, or JSON usage errors. Every outcome is still emitted as JSON.
Use the Agent tools
The built plugin is plugins/state-machine. It contains the bundled stdio MCP
server, manifest, and use-state-machine Skill. This repository does not modify
a personal marketplace or install itself automatically. Contributors can add a
local checkout for pre-release testing:
codex plugin marketplace add /absolute/path/to/state-machine
codex plugin add state-machine@state-machine
Restart the host and test in a new task after installing or upgrading.
Public tools:
| Tool | Result |
|---|---|
machine.validate |
strict structure and semantic graph diagnostics |
machine.step |
one accepted or rejected event with next snapshot and effects |
machine.simulate |
bounded event trace and final snapshot |
machine.find_path |
shortest structural route and required guard names |
machine.inspect |
compact states, events, transitions, statistics, and limits |
machine.diff |
bounded semantic changes between two valid specs |
All tools accept inline data, make no external changes, and publish strict input schemas. Results are checked against the executable output schemas before they leave the server. The six-tool discovery catalog is capped at 36 KiB by the runtime contract. A normal request should use one direct tool call.
Machine Spec
The canonical format is YAML or JSON with version: "0.1", an initial state,
declared context and event fields, named guards, and a state map. Each state/event
pair owns at most one transition. Context assignment uses only tagged literal,
event, or context value sources. Effects contain only a symbolic type and
input value sources.
See Machine Spec v0.1, the generated
machine-spec.schema.json, and the shipped
order.machine.yaml.
Verify the project
npm run check
This runs type checks, core tests, schema drift checks, all builds, built CLI and
MCP stdio tests, an isolated package install, browser interaction and responsive
regressions, and plugin/contract checks. Overall visual and business acceptance
remains the owner's rendered-runtime judgment; its current route is recorded in
docs/REVIEW_CONTRACT.md.
Contribution and release details are in CONTRIBUTING.md and
docs/RELEASE.md. Stable public identifiers are recorded in
docs/PRODUCT_IDENTITY.md.
Boundary
v0.1 intentionally excludes hierarchical, parallel, history, delayed, actor,
and eventless statechart semantics. It also excludes expression evaluation,
effect execution, production orchestration, AI-generated business rules, and
XState/SCXML compatibility claims. The current product boundary is maintained in
docs/PRODUCT_MODEL.md.
License
Step Switch is licensed under the Apache License 2.0. See LICENSE
and NOTICE. The standalone plugin includes its own copies plus the
license and attribution text for software bundled into the browser and MCP
distributions.
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。
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。
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。
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。
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。
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。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
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。