MCP Server

Nightward ships a stdio Model Context Protocol server so AI clients can scan, explain, compare, plan, preview Nightward's bounded actions, and request local approval for exact write tickets.

nw mcp serve

Most MCP tools are read-only. For writes, MCP can request a bounded action approval; the user approves or denies it locally in the Nightward CLI, TUI, or Raycast extension; then MCP can apply only that exact approved one-time ticket. MCP cannot accept the beta responsibility disclosure. That disclosure is Nightward's local one-time acknowledgement that write-capable beta actions are user-authorized, and MCP cannot self-confirm writes.

Client Setup

Install Nightward first, then add the same stdio command to your client. Use an absolute path such as /Users/alice/.local/bin/nw if the client does not inherit your shell PATH.

Config: Claude Code: claude mcp add, or Claude Desktop: claude_desktop_config.json

Use the CLI form for Claude Code. Use the JSON block in Claude Desktop's MCP config.

claude mcp add --transport stdio --scope user nightward -- nw mcp serve

# Claude Desktop JSON
{
  "mcpServers": {
    "nightward": {
      "type": "stdio",
      "command": "nw",
      "args": ["mcp", "serve"],
      "env": {}
    }
  }
}

Client references: Claude MCP, Cursor MCP, Codex MCP config, Windsurf Cascade MCP, and Google Antigravity's in-app "Manage MCP Servers -> View raw config" flow.

VS Code-style clients use a different key:

{
  "servers": {
    "nightward": {
      "type": "stdio",
      "command": "nw",
      "args": ["mcp", "serve"]
    }
  }
}

Restart or reload the AI client after editing its MCP config. A useful first prompt is: "Audit my AI setup with Nightward, explain the top risks, preview relevant actions, and request approval before applying any write action."

Tools

Tool	Purpose	Apply behavior
nightward_scan	Run a redacted HOME or workspace scan.	Read-only.
nightward_doctor	Return provider, schedule, disclosure, and action posture.	Read-only.
nightward_findings	List findings with severity/rule filters.	Read-only.
nightward_explain_finding	Return one finding by ID or unique prefix.	Read-only.
nightward_analysis	Run Nightward analysis with selected providers.	Read-only; online providers require opt-in.
nightward_explain_signal	Return one analysis signal by ID or prefix.	Read-only.
nightward_policy_check	Run the policy gate with optional analysis.	Read-only.
nightward_fix_plan	Generate plan-only remediation directions.	Read-only.
nightward_report_history	List saved scheduled reports.	Read-only.
nightward_report_changes	Compare two report files or the latest two saved reports.	Read-only.
nightward_actions_list	List bounded registry actions.	Read-only.
nightward_action_preview	Preview one registry action.	Read-only.
nightward_action_request	Request local approval for one exact registry action.	Writes Nightward approval state only.
nightward_action_status	Read one approval ticket status.	Read-only.
nightward_action_apply_approved	Apply one approved, unexpired, one-time ticket.	Destructive only after local approval.
nightward_rules	List rules and remediation metadata.	Read-only.
nightward_providers	List provider capabilities and status.	Read-only.

For casual AI-client use, prefer compact calls:

{
  "include_analysis": true,
  "compact": true,
  "limit": 25
}

Compact mode keeps pass/fail, threshold, summary counts, and bounded finding or signal metadata without flooding the model with every local detail.

Resources

• nightward://latest-summary
• nightward://latest-report
• nightward://rules
• nightward://providers
• nightward://schedule
• nightward://actions
• nightward://disclosure
• nightward://action-approvals
• nightward://report-history

Prompts

• audit_my_ai_setup
• explain_top_risks
• fix_this_finding
• set_up_providers
• compare_reports

These prompts are workflow starters for clients that expose MCP prompts. They tell the assistant to preview registry actions, request approval when a bounded action is useful, and only apply exact tickets already approved locally.

Safety Model

• Stdio only; no HTTP listener.
• No telemetry.
• No default network calls.
• Strict tool input schemas, server-side invalid-argument rejection, and structured output.
• Tool execution failures return isError: true, not protocol crashes.
• Online-capable providers stay blocked unless explicitly allowed.
• MCP cannot self-confirm local writes.
• MCP cannot accept the beta responsibility disclosure, Nightward's local one-time acknowledgement that write-capable beta actions are user-authorized.
• MCP approval requests record only Nightward-owned approval state. Applying an approved action consumes one exact one-time ticket and is audited.
• No arbitrary MCP/agent config mutation in MCP v1.
• Explicit workspace and report-diff paths must stay under NIGHTWARD_HOME, exist as the expected regular file or directory type, and avoid symlink components.
• Preview output is redacted and exposes write targets before any out-of-band apply.

Use nightward_action_preview before nightward_action_request. For provider installs that run a package manager such as Homebrew, Go, or npm, review the exact command, provider privacy boundary, and rollback expectations before approving the pending ticket in the CLI, TUI, or Raycast extension.

Registry Package

Nightward's existing npm launcher is the MCP Registry package target:

• npm package: @jsonbored/nightward
• MCP registry name: io.github.jsonbored/nightward
• metadata: server.json
• package verification: mcpName in packages/npm/package.json

The registry publishes metadata, not binaries. Nightward still ships platform-specific CLI archives through the normal release path: macOS arm64/amd64, Linux arm64/amd64, and Windows amd64. Windows ARM64 remains deferred. Windows schedule install remains preview/fallback only.