MCP server for AdsPower local API integration, enabling browser profile management, fingerprint configuration, proxy settings, and browser automation
AdsPower Local API MCP has 12 tools with visible schemas and descriptions, but suffers from pervasive gaps in parameter clarity, output documentation, and error handling. All tools have descriptions (good baseline), but most descriptions are under 50 characters (well below the 194-char baseline for A+ tools). Parameter descriptions are present but often generic (e.g., 'The browser ID to start' vs specifying format/range). Complex parameters like proxy_config and fingerprint_config are TypedDict-annotated in Python but lack JSON Schema enums for constrained fields. Output schemas are never documented, callers do not know what fields to expect in responses. Error responses are generic string templates ('Failed to X, error: {msg}') with no guidance on recovery steps. Tools like start_browser and create_browser have optional/required parameter ambiguities (e.g., browser_id vs serial_number both optional, group_id required but default='0' mentioned in description). No tool implements dry-run or confirmation patterns, critical for destructive tools like delete_browser. Naming is strong (verb_noun convention), but composition issues exist: close_browser and start_browser are clear, but move_browser(group_id) conflicts with conceptual clarity, is this moving the browser process or the profile config? No pagination documented for list tools (get_browser_list, get_group_list, get_application_list) despite no result-limit mention. Security: API keys/tokens not visible in source, but no explicit server-side injection pattern documented. Overall: solid foundation undermined by incomplete schemas, missing output docs, and lack of error recovery guidance.
Close a browser with the Browser ID.
Create a new browser profile.
Create a new browser group.
Delete a browser profile.
Get the list of all applications.
Get the list of all browser profiles.
Get the list of all groups.
Output schemas undocumented. Tools return string responses with no typed field documentation, forcing LLMs to parse free text. get_browser_list returns 'list of all browser profiles' but callers do not know if response contains browser_id, name, status, or creation_date.
List tools (get_browser_list, get_group_list, get_application_list) do not document pagination or result limits. No mention of how many items are returned, whether pagination is supported, or what to do with large result sets.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 47 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Get the list of currently opened browsers.
Move a browser to a different group.
Start a browser with the Browser ID or Serial Number. Must provide one of the two.
Update a browser profile.
Update a browser group.
Parameter descriptions are generic and lack constraint specifics. Example: 'The group id of the browser, must be a numeric string (e.g., "123")', the phrase 'e.g., "123"' is an example value that LLMs may reuse literally. Should be: 'Numeric string ID (1-9999). Use get_group_list to find valid IDs, or "0" for default group.' No enums for proxy_soft (expected: brightdata|brightauto|...|no_proxy), proxy_type, country codes, timezone, or browser versions.
Error responses are generic string templates with no recovery guidance. Example: start_browser returns 'Failed to start browser, error: {msg}'. Per pattern, errors should tell LLM what to do next: 'Browser ID not found. Call get_browser_list() to list available profiles, or use create_browser() to create a new one.'
delete_browser is a destructive tool with no dry-run or confirmation pattern. Agents can irreversibly delete profiles without a confirmation step. Per pattern, destructive ops should support confirm_before_execute or offer an undo companion tool.
Mutually exclusive parameters not documented. start_browser has both browser_id and serial_number as optional, with doc saying 'Must provide one of the two.' LLM may pass neither or both. Should explicitly state: 'Exactly one of browser_id or serial_number required; do not pass both.'
Complex nested config objects (proxy_config, fingerprint_config) are typed as TypedDict in Python but JSON Schema representation does not include nested field enums. Example: proxy_config.proxy_soft can be one of ~10 values, but LLM has no machine-readable constraint, must infer from description text.
Tool descriptions are under 50 characters (median baseline 194 chars). Examples: 'Get the list of all browser profiles.' (38 chars), 'Get the list of all groups.' (28 chars). Should include: what data structure, when to call it, and what the LLM can do next.