Skip to main content
Glama
DOC Model Safety Gate

DOC Model Safety Gate

by laceykes

DOC Model Safety Gate

An AI can draft the change. It cannot publish it.

The Safety Gate loop: an agent proposes a change, the gate holds it, a human reviews it, and only then does it publish

The DOC Model builds software in three layers — Data, Operations, Components. This MCP server enforces the DOC rule for AI agents working in the Operations layer: an agent can propose changes to your content, but every proposal is staged for human review, and only reviewed work can go live.

AI agents can write to your CMS the moment you connect them to it. This server puts a gate in between: everything an agent produces is staged for review, a human decides, and only reviewed work can go live. If an output is still pending, the publish call itself is refused — the gate is enforced by the backend, not by a warning label.

How the Safety Gate works: AI agent, Safety Gate, human review, published site — plus undo and a record of every decision

The rule

  1. Propose — a workflow runs in dry-run mode. Nothing can publish from a dry run.

  2. Stage — AI outputs land in the Safety Gate with their original and proposed values, field by field.

  3. Review — a human accepts, rejects, or edits each output. Bulk review is available for whole runs.

  4. Publish — a live run is only permitted once its review run has no pending or rejected outputs. Attempting to publish early returns an error: "Node … has pending AI output. Review all outputs before LIVE run."

  5. Roll back — a published run can be rolled back when snapshots exist.

Related MCP server: NavoCMS

Tools

Tool

What it does

listPendingReviews

List gate runs that still have outputs waiting for a decision.

getGateReview

Read one gate run's outputs: field, original value, AI value, status.

reviewOutput

Accept, reject, or edit a single output.

bulkReviewOutputs

Accept or reject every pending output in a run at once.

publishReviewedRun

Run a workflow live, gated on its review run. The review run ID is required — there is no way to call this "around" the gate.

rollbackRun

Roll a published workflow run back.

The server also includes workflow inspection and graph tools for the Safety Gate API; the gate tools above are the supported launch surface (see Status).

Quickstart

The server is a stdio MCP server that talks to a Safety Gate GraphQL API endpoint.

npx doc-model-safety-gate

Or from source:

npm install
npm run build
node dist/index.js

Configuration

Environment variable

Required

Purpose

DOCMODEL_API_ENDPOINT

yes

GraphQL endpoint of the Safety Gate API, e.g. http://localhost:4001/graphql

DOCMODEL_API_TOKEN

yes

Bearer token sent to the API

DOCMODEL_USER_ID

yes

Identity the server acts as (sent as the x-user-id header)

DOCMODEL_PROJECT_ID

for member access

Project the server is scoped to (x-project-id)

DOCMODEL_USER_ROLE

no

member (default) or admin

The older STRATA_* spellings of all five variables still work as fallback aliases.

Claude Desktop / Cursor

{
  "mcpServers": {
    "doc-model-safety-gate": {
      "command": "npx",
      "args": ["-y", "doc-model-safety-gate"],
      "env": {
        "DOCMODEL_API_ENDPOINT": "http://localhost:4001/graphql",
        "DOCMODEL_API_TOKEN": "your-token",
        "DOCMODEL_USER_ID": "your-user-id",
        "DOCMODEL_PROJECT_ID": "your-project-id",
        "DOCMODEL_USER_ROLE": "member"
      }
    }
  }
}

Claude Code

claude mcp add doc-model-safety-gate \
  -e DOCMODEL_API_ENDPOINT=http://localhost:4001/graphql \
  -e DOCMODEL_API_TOKEN=your-token \
  -e DOCMODEL_USER_ID=your-user-id \
  -e DOCMODEL_PROJECT_ID=your-project-id \
  -- npx -y doc-model-safety-gate

See it work

The loop at the top of this page is the real behavior, captured from a working local demo: an AI summary is proposed, staged in the Gate, and a publish attempted before review is refused by the backend; after the reviewer accepts, the identical call succeeds. The full build story — including the blocked publish, caught live — is told at thedocmodel.com.

The demo environment (Postgres + PostgREST + a mock CMS + the Safety Gate API) is not part of this repository; the server requires a Safety Gate API endpoint (see Status).

Status

Pre-1.0 and under active packaging. Verified end-to-end against a local Safety Gate API on 2026-11-12: pending reviews list and open correctly, human review flips output status, unreviewed publishes are refused by the backend, and reviewed publishes execute. The server requires a Safety Gate API endpoint; a hosted endpoint is not part of this release.

License

MIT — see LICENSE. Use it, build on it, ship it commercially if you like; the hosted layer and the DOC Model name stay ours.


Built by Lacey Kesler. The DOC Model Safety Gate is the enforcement layer of the DOC Model (Data · Operations · Components) — taught at thedocmodel.com: AI proposes, humans decide, and the audit trail shows both.

Available Tools

15 tools
bulkReviewOutputsA

Accept or reject every pending AI output in a run at once. Returns how many outputs were reviewed. Bulk review does not support edited; review edited outputs one at a time with reviewOutput.

Inputs: runId (required), action (required).

createWorkflowA

Create a workflow definition (the step-based editor path). The API requires at least one step: pass steps as an array of { position, stepType, configJson, isEnabled? }. For graph-based workflows, create the definition here, then add nodes and edges with the graph tools.

Inputs: name (required), steps (required), siteId (required), projectId (required), +3 more.

createWorkflowEdgeC

Connect two nodes with an edge

Inputs: workflowId (required), sourceNodeId (required), targetNodeId (required), condition.

createWorkflowNodeC

Add a node to a workflow graph

Inputs: nodeType (required), workflowId (required), label, config.

getGateReviewA

Get every AI output waiting in the Safety Gate for one run, field by field: the original value, the AI's proposed value, and its review status (pending, accepted, rejected, edited). This is the diff a human reviews before anything can go live.

Inputs: runId (required).

getNodeRunsB

Get node-level run results for a workflow run

Inputs: runId (required).

getRunStatusC

Get the status and details of a workflow run

Inputs: runId (required).

getWorkflowC

Get one workflow definition by ID.

Inputs: id (required).

listPendingReviewsA

List workflow runs whose AI outputs are waiting for human review in the Safety Gate, with a pending count for each. Nothing in these runs has been published.

Inputs: projectId (required).

listWorkflowNodesC

List all nodes in a workflow graph

Inputs: workflowId (required).

listWorkflowsA

List workflow definitions for a project, optionally filtered to one site.

Inputs: projectId (required), siteId.

publishReviewedRunA

Publish a reviewed run: runs the workflow LIVE using the outputs a human accepted or edited in the Safety Gate. The API refuses a LIVE run when any AI output is still pending or rejected, so an unreviewed change cannot be published through this tool.

Inputs: workflowId (required), reviewRunId (required).

reviewOutputA

Record a human review decision for one AI output in the Safety Gate: accepted, rejected, or edited. When the action is edited, pass editedValue with the corrected value. Reviewing never publishes anything by itself; a run can only go LIVE after every output is reviewed.

Inputs: id (required), action (required), editedValue.

rollbackRunC

Roll back a workflow run, reversing what it published.

Inputs: runId (required).

runWorkflowC

Run a workflow in DRY_RUN or LIVE mode

Inputs: mode (required), workflowId (required), reviewRunId.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.3/5.0

    Scored across 15 tools

    Disambiguation4/5

    Most tools map cleanly to a resource+action, and the review tools (listPendingReviews vs getGateReview, reviewOutput vs bulkReviewOutputs) are well separated by granularity. Minor overlap remains between runWorkflow in LIVE mode and publishReviewedRun, and between getRunStatus and getNodeRuns, where an agent must reason about which is authoritative.

    Naming Consistency5/5

    All 15 tools use camelCase with a verb-first pattern (list/get/create/run/review/publish/rollback + noun), including predictable variants like reviewOutput vs bulkReviewOutputs and listWorkflows vs listWorkflowNodes. No mixed conventions or stray vague verbs.

    Tool Count4/5

    15 tools is at the upper edge of the well-scoped band, but it covers two genuine sub-domains (workflow authoring/execution and the Safety Gate review lifecycle). Each tool has a defensible role, though the graph node/edge tools add weight for what may be an occasional editing path.

    Completeness3/5

    Read/create/run/review/publish/rollback are covered well, but the surface is asymmetric: workflows and graph nodes/edges can be created yet there is no updateWorkflow, deleteWorkflow, deleteWorkflowNode, or deleteWorkflowEdge, leaving dead ends in graph editing. There is also no way to list all runs for a workflow, only run-specific status lookups.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to draft, review, preview, and publish content across multiple websites through typed MCP tools. Ensures safe, policy-gated, and auditable publication without requiring a conventional CMS admin interface.
      Apache 2.0
    • A
      license
      B
      quality
      D
      maintenance
      Enables safe, human-in-the-loop AI document collaboration by letting agents create isolated workspaces, edit tracked files, open reviewable Change Requests, and get human approval before merging, with Git-backed conflict detection and resolution.
      35
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI agents to request human approval before consequential actions, with policy-based routing, phone-based approver decisions, and signed audit evidence.
      32 npm
      MIT