MCP server for reading and writing files to an Obsidian vault via HTTP API
This server has serious definition quality gaps. The `http_request` tool is a critical security risk. Tool names are reasonable (verb-noun format) but parameters are poorly documented. Output is unstructured dict returns with no schema documentation. No input validation, no actionable error messages, and no pagination support.
Add a new file to the Obsidian vault
Delete a file from the Obsidian vault
Make a custom HTTP request to any endpoint
Read a file from the Obsidian vault
Write content to an existing file in the Obsidian vault
ALL tool descriptions are missing or trivial. Every tool description ('Read a file from the Obsidian vault', 'Add a new file to the Obsidian vault', etc.) is only 30-50 chars but lacks actionable context. No guidance on WHEN to call each tool, WHY, or WHAT it returns. This violates pattern:tool-description.
Output schemas are completely undocumented. All tools return a Dict with 'status', 'headers', 'body', 'error' fields, but LLM has no way to know this. No structured output schema is visible in code.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 29 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 40 | - | v1 |
Parameter descriptions are minimal or missing. file_path is described as 'Path to the file in the vault' (36 chars) but does not specify: format constraints, allowed characters, max length, or whether relative paths are supported. 'method' in http_request is described as 'HTTP method (GET, POST, PUT, DELETE, etc.)' but does not list allowed enum values.
Critical security vulnerability: http_request tool allows agents to make arbitrary HTTP requests to any URL with any payload. No restrictions, no permission checks, no audit trail. This violates pattern:permission-gate and pattern:secret-injection. An agent can be tricked into exfiltrating data, hitting internal services, or bypassing rate limits. The tool should be removed or severely restricted (whitelist-only URLs, no arbitrary payloads).
No error handling guidance. vault_request() catches all exceptions and returns {'error': str(e), 'method': method, 'url': url} without categorizing errors as retryable, user-fixable, or fatal. If a file is not found (404), the LLM receives the raw HTTP status with no guidance on next steps. Per pattern:recovery-guide, errors must tell the LLM what to do next.
Destructive operations (delete_file) lack confirmation or dry-run support. Per pattern:confirmation-request, irreversible operations should support a dry-run or confirmation step. Agents can delete files without any validation or user confirmation, risking data loss.
No input validation. file_path is accepted as a raw string with no checks for path traversal attacks (e.g., '../../../etc/passwd'). 'method' in http_request is converted to uppercase but not validated against allowed HTTP methods. Per pattern:tool-gateway, all agent input must be sanitized.
Tool composition issue: read_file, add_file, write_file all return raw HTTP response dict with status/headers/body. An LLM calling read_file does not know whether it succeeded or failed without parsing the status code manually. There is no structured success/failure wrapper, no standard error format across tools. This violates pattern:response-shaper.
API credentials (VAULT_URL, API_TOKEN) are loaded from environment and hardcoded into DEFAULT_HEADERS. While not exposed as tool parameters (which is good), there is no mechanism for per-request credential scoping, no audit logging of which agent called what tool, and no rate limiting. Per pattern:audit-trail, agent-initiated actions must be traceable.
Tool overlap and naming ambiguity: add_file and write_file appear to do similar things (write content to a path). add_file uses POST, write_file uses PUT. The distinction is not explained in descriptions, so LLM must guess which to use. Per naming rubric: 'When multiple tools operate on the same resource, their names must make the distinction obvious.'