Observe legacy web UIs and expose them as dynamic MCP tools. Uses Playwright for UI observation, local Ollama LLM for element mapping, and exposes a REST API with WebSocket streaming for agent execution logs.
This HTTP-based MCP server exposes 11 tools for legacy UI automation via Playwright and LLM-based element mapping. Tool definitions are present but lack rigor: descriptions are verbose and narrative rather than LLM-optimized; input schemas exist but many parameters lack type constraints and validation guidance; output schemas are undocumented; error handling is minimal. The server conflates multiple concerns (observation, mapping, execution, CRUD) into single tools rather than composing atomic operations. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear write/read distinctions. Parameter naming is inconsistent (auth_credentials as object vs structured fields). No pagination, rate limiting, or result capping documented. Security concerns: credentials passed as parameters rather than server-side injection.
Register a new legacy application. Encrypts credentials before storing.
Execute a mapped MCP tool against the live legacy application.
Execute a tool action against the live legacy application. Locates the target element by bounding box or semantic intent and performs the specified action (click, fill, select, etc.).
WebSocket endpoint that streams agent_execution_logs updates in real time. Supports filtering by app_id and session_id.
Fetch the status and details of a tool execution.
Check the health status of the Dynamic Agentic Bridge service.
List all registered legacy applications.
Credentials passed as tool parameters (auth_credentials object in observe_legacy_ui, execute_tool_action, create_application). Secrets in parameters leak into agent logs and traces. Must use server-side secret injection via environment variables or vault.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear semantic distinctions. execute_tool_action and execute_tool are destructive writes; list_* and observe_* are read-only. Annotations enable safer agent planning and audit trails.
Output schemas undocumented. observe_legacy_ui returns ObservationResult (accessibility_tree, screenshot_b64, normalized_dom, state_hash, url, title, viewport_*) but this structure is not declared in tool definition. LLMs cannot plan downstream calls without knowing return fields.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2026-07-28+ | v2 |
List all mapped MCP tools for an application.
Takes an ObservationResult and calls a local Ollama model (moondream by default) with a screenshot and accessibility tree to identify interactive UI elements and their semantic intent. Returns validated MCPToolCandidate objects.
Launches a headless browser, navigates to a legacy application, and captures a structured snapshot of its UI state: accessibility tree, screenshot, and normalized DOM.
Trigger the observer + mapper pipeline for a given application. Runs as a background task — returns session_id immediately.
Parameter auth_credentials is an untyped object. No schema, no validation rules, no guidance on structure (form-based vs cookie-based vs header-based). LLMs cannot construct valid payloads. Should be enum-constrained or split into separate typed parameters.
No error handling guidance. Tools lack recovery hints (e.g., 'If navigation times out, check base_url and auth_credentials'). No categorization of errors as retryable vs user-fixable vs fatal. Agents cannot self-correct on failures.