An MCP server providing workspace management tools including log search, memory management, project planning, workspace querying, file mutations, and task execution with a focus on deterministic error handling and structured data access.
Alfred presents 11 tools with generally complete schemas and descriptions, but falls short of production-grade quality in several key areas. Naming is verb-driven and mostly clear (log_search, memory_put, plan_get), descriptions exist for all tools and most parameters, and pagination is implemented for listing/search operations. However, descriptions are often generic and lack LLM-optimized guidance on when to use each tool versus alternatives. Parameter descriptions are sparse or missing context on constraints and dependencies. Error handling is present but not all tools provide actionable recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear risk categorizations. Output schemas are partially documented but missing some chaining fields needed for agent composition. The codebase shows thoughtful implementation (locking for concurrent access, dry-run support for destructive ops, full-text search with ranking), but the tool interface prioritizes implementation completeness over agent usability.
Search through log files with filtering by level, source prefix, and text query. Supports pagination via cursor.
Delete a fact from memory. Supports dry-run mode to preview deletion without committing.
Retrieve a fact from memory by its ID.
List facts in memory with optional filtering by tags and sorting. Supports pagination.
Store or update a fact in memory with subject, content, category, citations, and tags. Preserves ID and updates mutable fields.
Full-text search over memory facts with ranking and filtering. Returns ranked matches with relevance scores.
Add a new item to the plan. Assigns the next sequential ID automatically. Acquires a lock to prevent concurrent modifications.
No tool annotations despite clear risk categorization. Tools are categorized by risk (READ_ONLY, WRITE, DESTRUCTIVE) in the evaluation metadata, but this information is not exposed in tool definitions via readOnlyHint, idempotentHint, or destructiveHint. Agents cannot infer safety constraints and may attempt retry patterns on non-idempotent operations.
Missing output schemas. While input schemas are well-defined, tool descriptions do not document return value structures. Agents cannot plan multi-step workflows confidently, they don't know if plan_add returns the assigned ID, what fields memory_get returns, or whether log_search results include timestamps for downstream filtering.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 55 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 41 | - | v1 |
Remove a plan item by ID. Acquires a lock to prevent concurrent modifications.
Replace an existing plan item with new content. Updates all fields except the ID. Acquires a lock to prevent concurrent modifications.
Retrieve all items from the project plan file (ProjectPlan.md). Returns parsed structured plan items.
Update the status of a plan item by ID. Cannot update other fields. Acquires a lock to prevent concurrent modifications.
Vague descriptions and missing decision guidance. Descriptions are present but often generic ('Store or update a fact', 'Remove a plan item'). No guidance on WHEN to use memory_put vs memory_search followed by edit, or plan_update vs plan_edit. LLMs must reason about these distinctions without explicit help, increasing error rates.
Overlapping tools without clear distinction. plan_update (status only) and plan_edit (all fields) both modify items, but the semantics differ. LLMs will waste reasoning cycles choosing between them. Similarly, memory_put (insert/update) and implicit memory_search + update pattern are ambiguous.
Pagination and result limits inconsistent. log_search and memory_search support pagination with cursor and limit, but no maximum result limit is documented (e.g., 'results capped at 1000'). memory_list and plan_get lack pagination support entirely, no cursor, no limit. Large result sets will bloat context.
Missing enum constraints on status fields. plan_update and plan_edit accept a 'status' parameter. The description lists valid values (e.g., 'planned', 'in-progress', 'completed', 'cancelled') but the schema does not declare them as an enum. Agents may pass invalid values and rely on error recovery instead of constraint validation.
Error handling lacks recovery guidance. Tests show error classification (e.g., 'InvalidArgument' for bad cursor), but tool descriptions do not include recovery hints. E.g., if memory_get returns 'not found', should the agent call memory_search first? If plan_update fails due to lock contention, should it retry or ask the user?
No dry-run support for destructive operations except memory_delete. plan_delete lacks a dry_run parameter, unlike memory_delete. This is inconsistent and increases the risk of accidental data loss when agents compose operations.