MCP server for deploying and managing AI-generated projects to live HTTPS URLs with hot-reload sandbox support, access control, and resource management
OpenBerth MCP provides 12 tools with complete input schemas and descriptions. Tool naming follows verb_noun conventions appropriately (deploy, update, status, list, destroy, etc.). Descriptions are moderately detailed (ranging 80-280 chars), providing context on when to use tools and operational guidance. However, several critical gaps reduce overall quality: (1) Output schemas are not documented in the provided source, descriptions mention what tools return but JSON Schema output specifications are absent; (2) Error handling guidance is minimal, tools indicate what errors can occur but don't guide LLM recovery paths; (3) Parameter descriptions lack specific format constraints and valid value ranges; (4) No tool annotations (readOnlyHint, destructiveHint) are visible in the code, despite clear risk classifications (READ_ONLY, WRITE, DESTRUCTIVE); (5) Some parameters are undescribed or generic. The server demonstrates solid foundational quality but lacks production-grade polish in output documentation and error taxonomy.
Deploy a small set of AI-generated files to a live HTTPS URL. Provide files as a map of filepath to content. Best for 1-100 files generated in conversation. For iterative development where you'll make multiple changes, use berth_sandbox_create instead — it supports instant hot-reload without full rebuilds. Framework is auto-detected. If detection fails, include a .berth.json file with override fields (language, build, start, install, dev).
Deploy a project directory from the local filesystem to a live HTTPS URL. Creates a tarball and uploads it. Respects .gitignore. PREFERRED over berth_deploy when: - The project already exists on disk (user said "deploy this", "deploy my app", referenced a path) - The project has more than 5 files - The project has dependencies (node_modules, venv, etc.) that shouldn't be sent inline For iterative development where you'll make multiple changes, use berth_sandbox_create instead.
Destroy a deployment and free its resources. This permanently removes the deployment, its container, data, and URL.
List active deployments. By default returns the caller's own deployments. Pass `all: true` to list every deployment on the server (read visibility is open — each entry includes `ownerId`/`ownerName`). Mutation still requires ownership or admin. Use this to find existing deployments before creating new ones. If the user wants to update an existing app, find it here first rather than creating a duplicate.
Output schemas not documented. Tools have detailed descriptions but no formal JSON Schema output specifications. LLMs cannot predict response structure or plan downstream operations.
Tool annotations (destructiveHint, readOnlyHint) not visible in source. Tools are classified with Risk labels (READ_ONLY, WRITE, DESTRUCTIVE) but not registered with MCP tool annotations, limiting agent safety and reasoning.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 63 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 22 | - | v1 |
Lock or unlock a deployment. A locked deployment keeps running and serving traffic, but rejects all code updates, metadata changes, access control changes, and destroy until unlocked. Use this to freeze a stable deployment and prevent accidental changes.
Get the container logs for a deployment. Useful for debugging build or runtime errors. ALWAYS check logs when: - A deployment status is 'failed' - The deployed app shows errors or blank pages - You need to debug application behavior Logs include both build output and runtime output.
Set access control on a deployment. Modes: 'basic_auth' (browser password prompt), 'api_key' (header auth), 'user' (require OpenBerth login), 'public' (remove protection). For basic_auth, provide username and password. For api_key, an API key is auto-generated if not provided — use it via the 'X-Api-Key' header. For 'user' mode, optionally provide a 'users' list to restrict access to specific usernames — if omitted, any authenticated user can access. Use 'public' to remove any existing protection.
Create a live development sandbox with hot reload. File changes via berth_sandbox_push apply instantly — no full rebuild needed. USE SANDBOX (not deploy) when: - The user wants to iterate on code (they'll make multiple changes) - You're building something step-by-step - You want to test changes quickly before finalizing Supports: static HTML, Node.js (Vite/Next.js/Nuxt/SvelteKit), Python (FastAPI/Flask/Django), Go. After creation, use berth_sandbox_push to update files instantly. When done iterating, use berth_sandbox_promote to create an optimized production deployment. Framework is auto-detected. If detection fails, include a .berth.json file with override fields (language, build, start, install, dev).
Get the source code files of a deployment. Returns all text files and their contents. Use this to inspect what code is running in a deployment. Requires the deployment ID. Use berth_update to modify and redeploy.
Get the status of a deployment (building, running, failed, etc.) and its URL. Use this after berth_deploy or berth_update to check if the build completed. Status values: 'building' (wait and check again), 'running' (ready to use), 'failed' (check berth_logs for errors).
Update an existing deployment with new inline files. Replaces all files and triggers a full rebuild. For iterative changes to a sandbox, use berth_sandbox_push — it's instant (no rebuild).
Update an existing deployment from a local project directory. Creates a tarball and uploads it. Respects .gitignore. PREFERRED over berth_update when the project exists on disk.
Error handling lacks recovery guidance. Descriptions note failure modes (e.g., 'failed' status, build errors) but do not guide LLM recovery, no actionable error messages, retry classification, or alternative paths documented.
Parameter constraints underspecified. Descriptions mention constraints (e.g., 'Max 100 files', 'default 512m') but lack formal validation rules, regexes, or enums. 'env' parameter accepts arbitrary objects with no documented structure.
protect_mode parameter lacks enum constraint. Descriptions list valid modes ('basic_auth', 'api_key', 'user', 'public') but no JSON Schema enum defined, inviting LLM hallucination of invalid modes.
No pagination documented. berth_list accepts an 'all' boolean but does not define limits, pagination tokens, or offset/limit parameters for large deployments.