Loading ToolBench
MCP protocol server for AI agent integration with proof-backed, drift-resistant code memory and context packing
AtlasMemory MCP server has 29 tools with consistent naming (verb_noun pattern) and basic schemas, but suffers from significant gaps in parameter descriptions, output schema documentation, and error handling guidance. Most tools lack detailed descriptions of what they return and when to use them. Parameter descriptions are minimal or absent for many inputs. No evidence of error recovery guidance or actionable error messages. Schemas are present but incomplete, many parameters lack type information or constraints. The server addresses a specialized domain (code memory/indexing) but the tool definitions do not meet production quality standards for LLM agent integration.
Add a new project memory entry.
Automatically refresh stale cards using LLM generation.
Generate compact project bootstrap capsule.
Build a context pack for a given objective
Delete a project memory entry.
Generate change-only capsule since last/git sha/timestamp.
Get valid anchor candidates for a file.
Get code health metrics for the project.
Output schemas not documented. Tools like build_task_pack, bootpack, deltapack, prove_claim return complex objects but no schema is visible describing the structure, fields, or types agents should expect. This forces LLMs to guess at response structure and breaks downstream tool chaining.
Parameter descriptions missing or minimal. Many parameters lack descriptions explaining their purpose, constraints, or expected format. E.g., 'max' in get_allowed_evidence, 'strict' in validate_file_card, 'proof' enum values across multiple tools lack explanation of what each mode does.
No error handling guidance. Tools lack descriptions of failure modes, recovery steps, or actionable error messages. E.g., index_file does not explain what happens if the path is invalid, or how to recover. Agents cannot self-correct without explicit error guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Get database statistics and coverage metrics.
Get files that a given file depends on.
Get files that depend on a given file.
Retrieve the file card for a specific file.
Analyze impact of changes to a file.
Get import statements from a file.
Generate a living project brief with architecture, key flows, and health metrics.
Retrieve project memory entries (context, gaps, priorities, milestones, learnings).
Get symbols (functions, classes, etc.) from a file.
Generate compact agent operating protocol.
Parse a single file, extract symbols/anchors, store them, and return extracted data.
Index the repository at the given path
List all indexed files with optional filtering.
Get or set the current project root path.
Find evidence anchors/snippets for a claim.
Batch prove multiple claims with dedup, registry cache, and proof budgeting.
Get a list of files that need card updates (missing or stale).
Search the indexed repository for relevant files
Update an existing project memory entry.
Store a validated FileCard.
Validate a FileCard JSON against schema and allowed anchors.
Destructive operations lack confirmation/dry-run support. delete_project_memory is marked DESTRUCTIVE but has no dry-run parameter or confirmation step. Agents can accidentally delete memory entries without safeguards.
Tool descriptions too brief. Most descriptions are 35 - 55 characters, below the 50 - 200 character baseline for LLM-optimized descriptions. E.g., 'Get valid anchor candidates for a file' (get_allowed_evidence) does not explain when to call it, what anchors are, or how to use the result.