mcp

MCP server combining PostgreSQL, Swagger/OpenAPI, and browser/frontend automation tools.

The Swagger side is designed as an execution layer for another agent:

• auth is managed inside MCP memory

• operation tools are generated from Swagger operationId

• consumers do not pass Authorization headers, API paths, or HTTP methods

The browser side is designed for frontend diagnostics and UI automation:

• stateful Playwright Chromium sessions

• screenshots and DOM/layout inspection

• frontend auth hydration via API login -> localStorage

• console and network diagnostics

Browser tool behavior

Registered browser tools:

• browser_open_session

• browser_close_session

• browser_navigate

• browser_auth_from_api_login

• browser_set_local_storage

• browser_seed_auth_state

• browser_open_profile_page

• browser_open_account_home

• browser_open_security_page

• browser_capture_profile_mobile

• browser_wait_for

• browser_click

• browser_fill

• browser_set_input_files

• browser_press

• browser_evaluate

• browser_get_text

• browser_get_attribute

• browser_screenshot

• browser_get_bounding_rect

• browser_get_computed_styles

• browser_assert_layout

• browser_save_storage_state

• browser_load_storage_state

• browser_get_console_logs

• browser_get_network_errors

• browser_inspect_page

Browser auth flow

Preferred frontend authentication flow:

1. Obtain API tokens using auth_login or browser_auth_from_api_login.

2. Write { accessToken, refreshToken, user: null } into frontend localStorage via browser_auth_from_api_login, browser_seed_auth_state, or browser_set_local_storage.

3. Navigate directly to the target authenticated route.

This avoids brittle UI-login dependencies and keeps browser automation stable.

Supported auth modes in high-level browser tools:

• apiLogin - perform API login and hydrate frontend auth state into localStorage

• useExistingMcpAuth - reuse an already authenticated MCP auth session (for example after auth_login)

• none - skip auth hydration entirely

Auth hardening notes:

• browser_auth_from_api_login and browser_inspect_page now return structured auth failures:

  • AUTH_API_LOGIN_FAILED

  • AUTH_REDIRECTED_TO_LOGIN

  • AUTH_SESSION_EXPIRED

• when a valid shared MCP auth session exists, browser auth can fall back to useExistingMcpAuth

• browser_set_local_storage works on an explicit origin and can optionally navigateToOrigin and/or reloadPage

• browser_seed_auth_state is a convenience wrapper for common auth objects like { accessToken, refreshToken, user }

• auth error payloads include:

  • finalUrl

  • debugScreenshotPath

  • debugHtmlPath

Browser file upload flow

browser_set_input_files supports:

• one or many files

• local disk paths

• in-memory files from text

• in-memory files from base64

• regular and multiple file inputs

Recommended input file descriptor shapes:

• local file path as plain string

• { "path": "/abs/file.png" }

• { "name": "avatar.txt", "mimeType": "text/plain", "text": "hello" }

• { "name": "avatar.png", "mimeType": "image/png", "base64": "..." }

When the page clears <input type="file"> inside its @change handler (common in Vue upload flows), the tool may return inputClearedAfterAttach: true and fall back to the requested file metadata for appliedFiles / fileCount instead of reporting a false 0.

Browser diagnostics

Browser diagnostics are intentionally split to reduce noise:

• consoleErrors - actual console errors and page errors

• consoleWarnings - warnings that survived filtering

• networkErrors - failed requests and HTTP 4xx/5xx responses

• ignoredNoiseCount - filtered Vite/dev noise or harmless net::ERR_ABORTED requests

Backward-compatible fields are still returned where practical:

• browser_get_console_logs still returns logs

• browser_get_network_errors still returns requests

Interpretation guidelines:

• non-zero ignoredNoiseCount is expected in dev-mode frontends

• investigate consoleErrors first, then networkErrors

• consoleWarnings are useful for regressions, but not all warnings are fatal

• filtered noise should not hide real app errors: only known Vite/dev chatter and harmless aborted requests are ignored

Artifacts are written under artifacts/browser/<sessionId>/....

Related MCP server: PostgreSQL MCP Server

QA helper tools

Registered QA tools (for browser/API end-to-end checks):

• qa_get_verification_code - latest registration verification code by email (requires PostgreSQL)

• qa_list_recent_users - list recent users by email prefix (requires PostgreSQL)

• qa_prepare_fixture_image - create a local JPEG in artifacts/qa/ for browser_set_input_files

• browser_scan_i18n_leaks - scan visible page text and document.title for untranslated i18n keys

Typical registration QA flow:

1. browser_open_session + /register

2. submit form in browser

3. qa_get_verification_code with the test email

4. fill OTP in browser

5. browser_scan_i18n_leaks on public/account pages

SQL/Swagger registration now retries startup connections (use DB_HOST=127.0.0.1 on WSL if localhost fails).

Vitour design reference tools

Vitour (VITOUR_ROOT, default ~/ide/vitour) is a static HTML template used for layout reference when building Outvento frontend screens. Vitour is served over HTTP (not file://).

Registered Vitour tools:

• vitour_list_pages - catalog of top-level *.html pages with Outvento mapping hints

• vitour_ensure_server - start or reuse python3 -m http.server for Vitour

• vitour_read_snippet - read HTML snippet from disk (optional CSS selector)

• vitour_open_page - open Playwright session on a Vitour page (requires browser tools)

• vitour_inspect_page - screenshot/styles/diagnostics for a Vitour page (requires browser tools)

Environment:

• VITOUR_ROOT - absolute path to the Vitour template directory

• VITOUR_BASE_URL - default http://127.0.0.1:8765

• VITOUR_STATIC_HOST - bind host, default 127.0.0.1

• VITOUR_STATIC_PORT - default 8765

• VITOUR_TOOLS_ENABLED - set false to disable the Vitour tool group

Typical workflow for news UI:

1. vitour_read_snippet with page: "blog" or page: "blog-details"

2. vitour_inspect_page with the same page slug for visual verification

3. Implement Vue pages in travelfront, then verify on http://outvento.test

See also vitour/AGENTS.md in the workspace.

Monitoring / latency report

Remote nginx latency report (same rt= metric as Grafana dashboard Outvento Response Time).

Registered monitoring tools:

• monitoring_latency_report - JSON report with avg/p50/p95/max and slowest requests for dev/prod API nginx

Implementation runs outvento/scripts/latency_report.sh over SSH (read-only: parses Docker nginx access logs).

Environment:

• OUTVENTO_ROOT - path to outvento repo (auto-detected as sibling ../outvento when present)

• MONITORING_SSH_HOST - default 167.86.76.119

• MONITORING_SSH_USER - default outvento

• MONITORING_SSH_KEY - default ~/.ssh/outvento_server

• MONITORING_TOOLS_ENABLED - set false to disable

Tool parameters:

• environment - dev, prod, or all (default all)

• since - docker logs window, e.g. 1h, 30m, 24h

• filter - human (exclude bots/scanners), api (/api/v1/ only), or all

• topN - number of slowest requests (default 15)

Fallback without MCP:

cd outvento && make latency-report
cd outvento && ENV=dev SINCE=1h make latency-report

Available in full and lite profiles when OUTVENTO_ROOT + SSH key exist.

Swagger tool behavior

Registered core Swagger tools:

• auth_login - stores token in MCP session memory

• auth_logout - clears stored token

• auth_status - shows auth session state

• call_api_raw - universal raw HTTP caller with custom method/path/query/headers/body/multipart control

• call_api_by_swagger - low-level operationId caller with internal auth header injection

• get_profile_page - high-level profile page wrapper

• get_translations - high-level translations wrapper

• update_profile - high-level profile update wrapper

• list_api_endpoints

• get_endpoint

• inspect_swagger_endpoint

• get_schema

• find_endpoint_by_keyword

call_api_raw contract

call_api_raw accepts:

• method

• either absolute url, or baseUrl + path

• pathParams

• query

• headers

• exactly one body mode:

  • jsonBody

  • textBody

  • multipart.fields[]

• timeoutMs

• expectedResponseType: auto | json | text | base64

• includeRawBody

• useExistingMcpAuth

multipart.fields[] item shape:

• name (required)

• one of:

  • value

  • filePath

  • text

  • base64

• optional fileName

• optional contentType

Response shape:

• ok

• request with resolved URL/method/body mode

• response with:

  • status

  • statusText

  • headers

  • contentType

  • body

  • rawBody (when requested or useful for debug)

  • parseError (when response could not be parsed as expected)

• structured error for 4xx/5xx, timeouts, or network failures

Swagger inspection helpers

inspect_swagger_endpoint accepts either:

• operationId, or

• path + method

Optional flags:

• includeSchemaDefinitions

• forceRefresh

It returns:

• resolved operationId

• combined required params

• request body content-types and schema refs

• response content-types and schema refs

• referenced schema names

• optional inline schemaDefinitions

Additionally, MCP generates one tool per Swagger operation:

• tool name pattern: api_<operation_id_normalized>

• fallback when operationId is missing: api_<method>_<path>

• input shape for generated tools:

  • pathParams (optional)

  • query (optional)

  • body (optional)

Example generated tools:

• api_get_profile

• api_update_profile

• api_get_translations

Wrapper mapping order is deterministic:

• fixed table tool_name -> operationId

• startup validation fails fast if mapped operationId is absent in Swagger

Authentication flow

Preferred usage:

1. Call auth_login once (or configure env credentials for auto-login).

2. Call wrapper tools or api_* tools mapped by Swagger operationId.

MCP handles:

• token storage in process memory

• automatic Authorization header injection

• clear unauthorized errors: Unauthorized: call auth_login first

Internal translations auth

Internal translation endpoints are handled separately from regular user/session auth:

• GET /api/v1/internal/translations/{locale}/{file}

• POST /api/v1/internal/translations/merge

For these endpoints MCP injects a dedicated internal bearer token automatically, even when Swagger does not declare security for the operation.

Behavior summary:

• generated tools like api_get_internal_translations and api_merge_internal_translations use the internal token automatically

• call_api_raw also auto-injects the same token for paths under /api/v1/internal/translations... when no explicit Authorization header is provided

• regular endpoints continue using the existing MCP auth session / auth_login flow unchanged

• when API_DEBUG=true, MCP logs:

  • whether an internal token source was found

  • whether an Authorization header was formed

  • only masked token metadata (len=<n> sha256=<prefix>), never the token value itself

Environment

• Create .env and/or .env.local in repository root.

• Env source priority is: process.env -> .env.local -> .env.

• SQL tools require: DB_HOST, DB_PORT, DB_USER, DB_PASSWORD, DB_NAME.

• Swagger tools require: SWAGGER_URL.

Auth/API controls:

• AUTH_LOGIN_PATH (default: /api/v1/login)

• AUTH_LOGIN_METHOD (default: post)

• AUTH_USERNAME_FIELD (default: login)

• AUTH_PASSWORD_FIELD (default: password)

• AUTH_TOKEN_FIELD_PATH (default: content.accessToken)

• AUTH_REFRESH_TOKEN_FIELD_PATH (default: content.refreshToken)

• AUTH_TOKEN_TYPE_FIELD_PATH (default: content.tokenType)

• AUTH_DEFAULT_TOKEN_TYPE (default: Bearer)

• AUTH_TOKEN (optional pre-seeded static token)

• AUTH_REFRESH_TOKEN (optional pre-seeded static refresh token)

• AUTH_USERNAME / AUTH_PASSWORD (optional for auto-login)

• AUTH_AUTO_LOGIN (default: true)

• API_DEBUG (default: false) - enables masked debug logging for auth/header diagnostics

• API_REQUEST_TIMEOUT_MS (default: 15000)

• API_RETRY_ON_UNAUTHORIZED (default: true)

• TLS / HTTPS controls for Swagger, auth, and raw API calls:

  • API_TLS_REJECT_UNAUTHORIZED (default: true)

  • API_TLS_INSECURE_SKIP_VERIFY (default: false, alias fallback when API_TLS_REJECT_UNAUTHORIZED is not set)

  • API_TLS_CA_CERT_PATH (optional absolute path to a PEM root/intermediate CA bundle)

  • backward-compatible aliases are also accepted for Swagger-specific migrations:

    • SWAGGER_TLS_REJECT_UNAUTHORIZED

    • SWAGGER_TLS_INSECURE_SKIP_VERIFY

    • SWAGGER_TLS_CA_CERT_PATH

Internal API token controls (/api/v1/internal/translations, /api/v1/internal/news, …):

• token priority:

  1. INTERNAL_API_TOKEN

  2. AUTH_TOKEN (legacy fallback only)

• token type priority:

  1. INTERNAL_API_TOKEN_TYPE

  2. AUTH_DEFAULT_TOKEN_TYPE

• default token type is Bearer

Recommended configuration:

INTERNAL_API_TOKEN=your-internal-token
INTERNAL_API_TOKEN_TYPE=Bearer
API_DEBUG=false

For internal/self-signed HTTPS environments, prefer trusting the CA instead of disabling verification:

SWAGGER_URL=https://api.outvento.test/docs?api-docs.json
API_TLS_CA_CERT_PATH=/abs/path/to/root-ca.pem

Temporary diagnostic-only fallback:

SWAGGER_URL=https://api.outvento.test/docs?api-docs.json
API_TLS_INSECURE_SKIP_VERIFY=true

API_TLS_INSECURE_SKIP_VERIFY=true should only be used for short-lived local diagnostics when the proper CA chain is not yet available.

Browser controls:

• BROWSER_TOOLS_ENABLED (default: true)

• BROWSER_HEADLESS_DEFAULT (default: true)

• BROWSER_SESSION_TTL_MS (default: 600000)

• BROWSER_CLEANUP_INTERVAL_MS (default: 60000)

• BROWSER_NAVIGATION_TIMEOUT_MS (default: 30000)

• BROWSER_ACTION_TIMEOUT_MS (default: 30000)

• BROWSER_MAX_CONSOLE_ENTRIES (default: 200)

• BROWSER_MAX_NETWORK_ERRORS (default: 200)

• FRONTEND_AUTH_STORAGE_KEY (default: auth)

• BROWSER_ARTIFACTS_DIR (default: artifacts/browser)

If required variables are missing, corresponding tool groups are not registered. tool_status is always available.

Swagger generation controls:

• SWAGGER_GENERATED_API_TOOLS_ENABLED (default: true) - register one api_<operationId> tool per Swagger operation.

• When false, keep meta/auth/call tools only (call_api_by_swagger, call_api_raw, inspect_swagger_endpoint, ...).

• Use false for IDEs with a low MCP tool limit (for example PHPStorm GitHub Copilot: 128 tools).

Lite profile (PHPStorm / Copilot)

Full registration exposes ~206 tools. Lite mode keeps ~18:

Start lite server:

npm run start:lite

Equivalent .env.local flags (see .env.lite.example):

SWAGGER_GENERATED_API_TOOLS_ENABLED=false
BROWSER_TOOLS_ENABLED=false
VITOUR_TOOLS_ENABLED=false

Reference MCP config for JetBrains: .ai/mcp/mcp.lite.json (outvento-mcp-lite server name).

Windows PHPStorm global config path:

C:\Users\Allximik\AppData\Local\github-copilot\intellij\mcp.json

JetBrains Copilot uses root key servers (not Cursor's mcpServers).

PHPStorm setup

1. Keep Cursor on full MCP: npm start (206 tools).

2. In PHPStorm, use lite config at ...\github-copilot\intellij\mcp.json (see .ai/mcp/mcp.lite.json).

   • Or open via Settings → Tools → AI Assistant → Model Context Protocol (MCP)
     (or Settings → Plugins → GitHub Copilot → MCP, depending on version)

3. Enable only outvento-mcp-lite in PHPStorm. Do not attach the full outvento-mcp server there.

4. Restart PHPStorm after editing mcp.json (Copilot reloads MCP config on IDE restart).

5. After start, verify with tool_status → expect registeredTools: "17/17" (or 19/19 if deterministic wrappers differ).

6. Call APIs via call_api_by_swagger with operationId (for example login, profilePage) instead of api_login.

If PHPStorm runs inside WSL natively, replace wsl.exe ... with:

bash -lc "cd /home/allximik/ide/outvento-mcp && npm run start:lite"

Structure

• index.js - bootstrap and wiring.

• src/config/env.js - environment config parsing.

• src/infrastructure/db/client.js - PostgreSQL client factory.

• src/services/auth/auth-session.js - auth session and token management.

• src/services/browser/session-manager.js - stateful Playwright browser/context/page lifecycle.

• src/services/browser/playwright-service.js - navigate, wait, screenshot, DOM geometry, computed styles.

• src/services/browser/artifact-service.js - runtime screenshots, HTML dumps, JSON artifacts.

• src/services/browser/auth-bridge.js - API login to frontend localStorage auth hydration.

• src/services/swagger/swagger-cache.js - Swagger loading and cache.

• src/services/swagger/http-executor.js - shared raw HTTP execution, query/path resolution, multipart handling, and response parsing.

• src/tools/register-browser-tools.js - browser automation MCP tools.

• src/tools/register-sql-tools.js - SQL MCP tools.

• src/tools/register-swagger-tools.js - auth and generated Swagger operation tools.

• src/tools/register-health-tool.js - combined health check tool.

• src/tools/register-status-tool.js - tool availability diagnostics.

• scripts/browser-smoke.js - local end-to-end smoke scenario for browser tools.

• scripts/swagger-smoke.js - local smoke scenario for raw API and Swagger inspection tools.

Run

npm install
npx playwright install chromium
npm start          # full profile (~205 tools, Cursor)
npm run start:lite # lite profile (~17 tools, PHPStorm Copilot)

Quick check

npm run check
npm run smoke:swagger
npm run smoke:browser

Example raw API tool calls

GET with query params and custom headers

{
  "method": "get",
  "baseUrl": "http://outvento.test",
  "path": "/api/v1/users/{userId}",
  "pathParams": {
    "userId": 42
  },
  "query": {
    "details": "full"
  },
  "headers": {
    "Authorization": "Bearer <token>",
    "X-Locale": "uk",
    "X-Requested-With": "XMLHttpRequest"
  },
  "includeRawBody": true
}

POST JSON

{
  "method": "post",
  "baseUrl": "http://outvento.test",
  "path": "/api/v1/profile",
  "headers": {
    "X-Locale": "en"
  },
  "jsonBody": {
    "firstName": "Ada",
    "lastName": "Lovelace"
  }
}

POST multipart/form-data

{
  "method": "post",
  "baseUrl": "http://outvento.test",
  "path": "/api/v1/user/avatar",
  "useExistingMcpAuth": true,
  "multipart": {
    "fields": [
      {
        "name": "avatar",
        "filePath": "/abs/path/avatar.png",
        "contentType": "image/png"
      },
      {
        "name": "folder",
        "value": "profile"
      }
    ]
  }
}

Parse non-JSON responses for debug

{
  "method": "get",
  "url": "http://outvento.test/api/v1/debug/raw-response",
  "expectedResponseType": "json",
  "includeRawBody": true
}

Example browser tool calls

Open a mobile session

{
  "baseUrl": "http://outvento.test",
  "device": "iPhone 14 Pro Max",
  "headless": true
}

Authenticate frontend via API login + localStorage

{
  "sessionId": "browser-session-001",
  "baseUrl": "http://outvento.test",
  "login": "user@example.com",
  "password": "secret"
}

Set arbitrary localStorage state on an origin

{
  "sessionId": "browser-session-001",
  "origin": "http://outvento.test",
  "key": "settings",
  "value": {
    "locale": "uk",
    "theme": "dark"
  },
  "navigateToOrigin": true,
  "reloadPage": true
}

Seed auth state directly

{
  "sessionId": "browser-session-001",
  "origin": "http://outvento.test",
  "accessToken": "<access-token>",
  "refreshToken": "<refresh-token>",
  "user": {
    "id": 1,
    "email": "user@example.com"
  }
}

Upload a single file from text

{
  "sessionId": "browser-session-001",
  "selector": "input[type='file']",
  "files": [
    {
      "name": "avatar.txt",
      "mimeType": "text/plain",
      "text": "avatar-content"
    }
  ]
}

Upload multiple files to <input multiple>

{
  "sessionId": "browser-session-001",
  "selector": "input[type='file'][multiple]",
  "files": [
    "/abs/path/avatar.png",
    {
      "name": "meta.json",
      "mimeType": "application/json",
      "base64": "eyJoZWxsbyI6IndvcmxkIn0="
    }
  ]
}

Example Swagger inspection calls

Inspect by path + method

{
  "path": "/api/v1/user/avatar",
  "method": "post",
  "includeSchemaDefinitions": true,
  "forceRefresh": true
}

Inspect by operationId

{
  "operationId": "saveUserAvatar",
  "includeSchemaDefinitions": true
}

Open profile page with existing MCP auth

{
  "baseUrl": "http://outvento.test",
  "auth": {
    "mode": "useExistingMcpAuth"
  }
}

Open security page in a dedicated session

{
  "baseUrl": "http://outvento.test",
  "device": "Desktop Chrome",
  "auth": {
    "mode": "apiLogin",
    "login": "user@example.com",
    "password": "secret"
  }
}

Navigate and inspect

{
  "sessionId": "browser-session-001",
  "url": "http://outvento.test/account/profile",
  "waitUntil": "networkidle"
}

Interact with the page

{
  "sessionId": "browser-session-001",
  "selector": ".profile-input",
  "value": "Ada Lovelace"
}

{
  "sessionId": "browser-session-001",
  "selector": ".save-button"
}

{
  "sessionId": "browser-session-001",
  "selector": ".shortcut-input",
  "key": "Enter"
}

{
  "sessionId": "browser-session-001",
  "expression": "document.querySelector('.save-status').textContent"
}

{
  "sessionId": "browser-session-001",
  "selector": ".save-status"
}

{
  "sessionId": "browser-session-001",
  "selector": ".save-button",
  "name": "data-role"
}

{
  "sessionId": "browser-session-001",
  "selector": ".profile-page",
  "assertions": {
    "topLessThanOrEqual": 80,
    "widthGreaterThanOrEqual": 390
  }
}

Save and restore storage state

{
  "sessionId": "browser-session-001",
  "fileName": "auth-state.json"
}

{
  "sessionId": "browser-session-002",
  "path": "/.../artifacts/browser/browser-session-001/auth-state.json"
}

{
  "sessionId": "browser-session-001",
  "selector": ".profile-page",
  "includeParents": true,
  "stopAt": "body"
}

High-level inspection

{
  "baseUrl": "http://outvento.test",
  "url": "http://outvento.test/account/profile",
  "device": "iPhone 14 Pro Max",
  "headless": true,
  "auth": {
    "mode": "apiLogin",
    "login": "user@example.com",
    "password": "secret"
  },
  "targetSelector": ".profile-page",
  "captureStyles": true,
  "captureConsole": true,
  "captureNetworkErrors": true,
  "takeFullPageScreenshot": true,
  "takeElementScreenshot": true
}

Mobile profile capture

{
  "baseUrl": "http://outvento.test",
  "headless": true,
  "auth": {
    "mode": "useExistingMcpAuth"
  }
}

Example smoke flow

Short end-to-end browser smoke flow:

1. auth_login (optional but recommended when useExistingMcpAuth is used)

2. browser_open_profile_page

3. browser_assert_layout

4. browser_get_console_logs

5. browser_get_network_errors

6. browser_open_security_page or browser_open_account_home

7. browser_capture_profile_mobile

8. browser_close_session

Browser smoke scenario

The included smoke harness starts a local mock frontend/API and validates this flow:

1. browser_open_session

2. browser_auth_from_api_login

3. browser_open_profile_page

4. browser_open_account_home

5. browser_open_security_page

6. browser_capture_profile_mobile

7. browser_navigate

8. browser_wait_for

9. browser_fill

10. browser_set_input_files (single file)

11. browser_set_input_files (multiple files)

12. browser_click

13. browser_press

14. browser_evaluate

15. browser_get_text

16. browser_get_attribute

17. browser_assert_layout

18. browser_save_storage_state

19. browser_load_storage_state

20. browser_set_local_storage

21. browser_seed_auth_state

22. browser_screenshot (full page)

23. browser_screenshot (element)

24. browser_get_bounding_rect

25. browser_get_computed_styles

26. browser_get_console_logs

27. browser_get_network_errors

28. browser_inspect_page

29. auth failure coverage for AUTH_API_LOGIN_FAILED, AUTH_REDIRECTED_TO_LOGIN, AUTH_SESSION_EXPIRED

30. browser_close_session

Artifacts are written under artifacts/browser/<sessionId>/... and are intentionally ignored by Git.

Agent runbook (autonomous mode)

Use this policy for any AI agent that consumes this MCP server.

1) Mandatory first-step diagnostics

Always run, in this order:

1. tool_status

2. health

3. auth_status

Do not ask the user what to do before these checks are complete.

2) Decision policy

• tool_status=ok and health=ok:

  • server is healthy

  • if request requires auth and auth_status is not authenticated, run auth_login (when credentials are available) or ask only for missing credentials

• tool_status=ok and health!=ok:

  • perform a soft recovery:

    1. auth_logout

    2. re-run auth_status

    3. re-run health

  • if still unhealthy, escalate to hard restart instructions

• tool_status!=ok:

  • skip soft recovery and escalate directly to hard restart instructions

3) Soft recovery definition

Soft recovery means MCP session reset only (token/session cleanup), not process restart:

• auth_logout

• re-check tool_status, health, auth_status

4) Hard restart policy

If hard restart is required, the agent must explicitly state that host-level restart is outside MCP tool scope. Then provide concrete commands by runtime option:

# Docker Compose
docker-compose restart <mcp_service_name>
docker-compose restart

# systemd
sudo systemctl restart mcp

If runtime is unknown, ask one short question: docker, systemd, or other?

5) Response format requirements

Every operational response should include:

1. Current state (tool_status, health, auth_status)

2. Action taken automatically

3. Next automatic step (or one minimal blocking question)

Avoid open-ended prompts like "How do we proceed?" when a deterministic next step exists.