Stepgate

Agents can't skip steps. Write an agent's procedure once, as a YAML stepfile, and every step is gated: the agent moves on only when a mechanical check passes, never on its own word.

One file runs anywhere. A stepfile has no packages, no versions to pin and no code to deploy, so it moves as a single file to any MCP client, such as Claude Code, Cursor or an agent you wrote, and runs on the model that client already uses.

A stepfile declares its inputs, the remote APIs and MCP servers it may call, and an ordered list of steps. Each step says what output it must produce and which gates check that output. The file names no model and no framework.

Stepgate runs stepfiles. It is an MCP server that offers each stepfile as a tool. When a client's agent calls it, Stepgate shows the agent one step at a time, makes every API call the step needs, and moves on only when the step's gates pass.

Make the procedure you already wrote enforceable

A skill or runbook written as markdown tells an agent what to do, and the agent decides how much of it to follow. Here is one:

---
name: package-notes
description: Writes an upgrade note for each npm package, with versions taken from the registry.
---
1. Look up each package's latest version on the npm registry.
2. Write a one-line upgrade note for each. Never state a version the registry did not return.

stepgate_outline turns it into a skeleton, and the finished stepfile makes both lines binding. Stepgate does the lookup itself, so the agent never fetches or copies a version. The agent only writes the notes, and a gate rejects any note whose version differs from what the registry returned, naming the rows that broke it:

stepgate: "1"
id: package-notes
description: Writes an upgrade note for each npm package, with versions taken from the registry.
inputs:
  type: object
  required: [packages]
  properties:
    packages: { type: array, minItems: 1, maxItems: 20, items: { type: string, pattern: "^[a-z0-9][a-z0-9._-]*$" } }
tools:
  npm:
    openapi:
      server: https://registry.npmjs.org
      document:
        openapi: 3.1.0
        info: { title: npm registry, version: "1" }
        paths:
          /{name}/latest:
            get:
              operationId: getLatest
              parameters: [{ name: name, in: path, required: true, schema: { type: string } }]
    exposes: [getLatest]
steps:
  - id: look-up
    do:
      calls:
        - id: latest
          operation: getLatest
          each: { var: inputs.packages }
          arguments: { name: { var: item } }
      output:
        packages: { map: [{ var: responses.latest }, { object: [[name, { var: name }], [latest, { var: version }]] }] }
    produces:
      type: object
      required: [packages]
      properties: { packages: { type: array } }
  - id: notes
    instructions: |
      Write a one-line upgrade note for each of these packages, saying what
      its latest version is and whether that is a new major version:

      {{steps.look-up.packages}}
    produces:
      type: object
      required: [notes]
      properties:
        notes:
          type: array
          items:
            type: object
            required: [name, latest, note]
            properties: { name: { type: string }, latest: { type: string }, note: { type: string } }
    gates:
      - id: versions-from-registry
        message: Every note must give the version the registry returned for its package.
        predicate:
          none:
            - join: [{ var: output.notes }, { var: steps.look-up.packages }, name, name]
            - or: [{ "==": [{ var: right }, null] }, { "!=": [{ var: left.latest }, { var: right.latest }] }]
    retries: 2

Why

When an agent is handed a plan as text, it decides how much of the plan to follow, a step counts as done when the agent says so, and nothing records afterwards what actually ran. A stepfile moves those decisions out of the model:

• Steps run in order, one at a time. The agent is shown only the current step's instructions and operations, never a later step, so it cannot skip ahead. Earlier steps stay in its own conversation.
• Gates decide, not the model. A step passes only when its output satisfies JSON Schema, JSONLogic or an HTTP verifier, and gates can check that output against what the APIs actually returned, so a fabricated value fails. A failed gate's diagnosis goes back to the model for a bounded number of retries.
• The model judges, the server computes. A mechanical step makes its API calls and builds its output from a template, with no model involved, and a derived field fills in a count or a lookup after the model submits. The model is left the work that needs judgement.
• Writes wait for a person. An approve gate asks a person to confirm a step's output through an MCP elicitation before a later mechanical step writes it, so what reaches Jira is what the person saw (which clients show the form).
• The model never holds a key. The server makes every tool call and attaches credentials itself, and it refuses requests to hosts the stepfile does not declare.
• Every run leaves a record. A hash-chained ledger lists each step, tool call, gate verdict and retry, and editing it afterwards breaks the chain, which stepgate --verify detects.
• Nothing to install on the client side. Stepfiles call remote APIs only, and the client adds one MCP server to its configuration. Stepgate needs no model key: the client's own model does the reasoning, and the same file gives the same path through its steps whichever model that is.

Quick start

Add the server to your MCP client's configuration, naming one or more stepfiles from the catalog, or giving absolute paths to your own .stepfile.yaml files (details):

{
  "mcpServers": {
    "stepgate": {
      "command": "npx",
      "args": ["-y", "stepgate", "market-research"],
      "env": { "TAVILY_API_KEY": "tvly-..." }
    }
  }
}

The client sees a market-research tool. Ask your agent to run it for { "brand": "Oatly", "market": "UK plant-based milk" } and it works through four steps (search, filter, analyse, report) with stepgate_call and stepgate_submit, ending with every step's output. docs/connect.md has the configuration for Claude Code, Claude Desktop, Cursor, VS Code and other clients.

Every server also offers tools for writing stepfiles: ask your agent to write one for your use case, and it can read the format, inspect the APIs, validate its draft and try it through Stepgate (details).

To serve over HTTP instead of stdio, run npx -y stepgate --http 3100 market-research and connect to http://127.0.0.1:3100/mcp. npx -y stepgate --list shows the catalog, and --help lists every option, including these:

• --watch reloads your stepfiles when you save them, while you write one.
• --test <stepfile> checks its gates offline against recorded cases, and --record-cases <dir> records those cases from a real run.
• --auth <stepfile> <credential> signs in to an MCP server that uses MCP authorization and prints the variables to set (details).
• --verify <ledger> checks a run's ledger for edits.

A stepfile

stepgate: "1"
id: market-research
inputs:
  type: object
  required: [brand, market]
  properties:
    brand: { type: string }
    market: { type: string }

credentials:
  tavily:
    kind: bearer
    hosts: [mcp.tavily.com]
    description: Web search, used only by the search step.

tools:
  tavily:
    mcp: { url: "https://mcp.tavily.com/mcp/" }
    credential: tavily
    exposes: [{ name: tavily_search, effect: read }]

steps:
  - id: search
    tools: [tavily_search]
    instructions: |
      Run at least six searches about {{inputs.brand}} in {{inputs.market}}.
      Submit every result as a source with an id of the form S-01.
    produces:
      type: object
      required: [sources]
      properties:
        sources: { type: array }
    gates:
      - id: enough-sources
        schema: { properties: { sources: { minItems: 12 } } }
      - id: domain-breadth
        message: Sources must span at least six distinct domains.
        predicate:
          ">=":
            - { length: { unique: { map: [{ var: output.sources }, { host: { var: url } }] } } }
            - 6
    retries: 2
  # ... filter, analyse and report steps

Credentials say what is needed, never where it lives: the server reads tavily from TAVILY_API_KEY. effect: read marks the search as safe to retry. The complete file is stepfiles/marketing/market-research, and editors that support yaml-language-server validate against server/schema/stepfile.schema.json, which the npm package also ships.

Catalog

stepfiles/ is a community catalog of stepfiles, reviewed and shipped with the npm package, so each one runs by name. Built something repeatable? Adding it is one folder and one pull request: see stepfiles/README.md, or suggest an idea.

Documentation

• docs/stepfile.md: how to write a stepfile: fields, tools, credentials, steps and gates.
• docs/connect.md: connecting Claude Code, Claude Desktop and other MCP clients.
• docs/how-it-works.md: what Stepgate does during a run, its limits, errors and ledger.
• server/schema/stepfile.schema.json: the JSON Schema for stepfiles.
• CONTEXT.md: the project's vocabulary.

Contributing

See CONTRIBUTING.md. To report a vulnerability, follow SECURITY.md. Everyone taking part is expected to follow the Code of Conduct.

License

Apache License 2.0