Loading ToolBench
Context-sync protocol for AI agents. Every agent has a mailbox. Multi-agent coordination with thread-based messaging, context graphs, and codebase indexing.
AgentsMCP defines 29 tools with reasonable naming (verb_noun pattern) and descriptions, but exhibits critical gaps in schema completeness and parameter documentation. Most tools lack visible input schema definitions in the source code provided. Descriptions are present but often generic (e.g., 'Get unread messages' without guidance on when to use vs. similar tools). Parameter descriptions exist but are sparse, many lack type constraints, ranges, or format guidance. No output schemas are documented. Error handling is minimal; tools do not guide recovery or categorize errors. The server is STDIO-only, which is a hard transport limitation. Overall, this is a D-grade server with foundational issues that would require significant rework for production use.
Connect two graph nodes with a typed, directed edge. Edge types: references (message→file), contains (file→symbol), resolves (symbol→task), depends_on (symbol→symbol), semantic (any→any).
Analyze the impact of changes on legacy mainframe systems. Returns a task ID for polling.
Get a briefing of recent context relevant to a task: recent threads, graph nodes, and index entries. Useful at the start of a session to orient yourself.
Delete a memory entry by ID.
Get neighboring chunks in the vector store for a given chunk ID. Useful for exploring semantic relationships.
Look up a specific codebase index entry by key. Use this INSTEAD of reading an entire file when you just need to know what a file does, what it exports, or its role in the architecture.
No visible input schemas in source code. Tools declare parameters (to, payload, threadId, etc.) but JSON Schema definitions are not shown. Cannot verify type constraints, required fields, or format specifications.
No output schemas documented. Tools return data (threads, graph nodes, index entries, memories) but response structure is not formally specified. LLMs cannot plan downstream tool calls or extract required fields reliably.
Minimal error handling guidance. Tools do not indicate which errors are retryable, user-fixable, or fatal. No recovery hints provided (e.g., 'If thread not found, call agentsmcp_threads() to list available threads').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2026-07-28+ | v2 |
Check the health and status of the agent's memory and storage systems.
Trigger an async improvement pass on the agent's knowledge base (deduplication, compression, rule extraction). Returns a task ID for polling.
Extract and store learned rules from recent interactions. Rules are patterns that can guide future decisions.
List all stored data (memories, rules, graph nodes, index entries) with optional filtering.
Mark a thread as read for this agent.
Parse COBOL source code and extract semantic nodes (paragraphs, sections, variables, copybooks). Returns an AST-like structure suitable for graph ingestion.
Parse JCL (Job Control Language) source code and extract semantic nodes (job steps, datasets, procedures). Returns an AST-like structure suitable for graph ingestion.
Parse mainframe source code (COBOL, JCL, PL/I, REXX) with automatic language detection. Returns semantic nodes and relationships.
List visible participants on a thread with their roles (to/cc/bcc). BCC participants are only shown if this agent bcc'd them.
Remove stale or low-confidence data (old memories, unused nodes, outdated index entries).
Search the context graph by keywords and return matching nodes plus all nodes reachable within 2 hops. Use this INSTEAD of grepping and reading files — it returns structured context (files, symbols, decisions, tasks) with their relationships.
Retrieve memories matching a query. Returns ranked results with relevance scores.
Get unread messages addressed to this agent, with full thread context attached to each. Use this at the start of a turn to pick up cold. Pass recent=3 (default) to keep context compact; increase only when you need deeper history.
Store a memory entry (fact, insight, or learned pattern) in the agent's long-term memory. Memories are indexed and retrievable via agentsmcp_recall.
Reply to every visible participant on a thread (excluding the sender and BCC'd agents).
Search the codebase index by keywords and optional category. Returns summaries ranked by relevance.
Send a message to another agent. Auto-creates a thread if none exists between sender and recipient. Use cc for active participants, bcc for silent ones. contextSnapshot captures your current state so the recipient can pick up cold.
Rejoin a thread with full assembled context (snapshot + recent 10 messages verbatim + summary of older ones). Use after a restart or when picking up a stale thread.
Poll the status of a background task by ID. Returns status (pending, running, completed, failed) and result if available.
List all threads this agent is part of.
List unread context frames without consuming them.
Register a codebase index entry (file summary, symbol summary, API contract, config description, or architecture note). Call this when you finish working on a file to persist a ~200-token summary so future agents can look it up instead of reading the full file.
Register a context graph node (file, symbol, decision, task, or message). Call this when you create/modify files, implement symbols, make design decisions, or track tasks. The node persists across sessions so future agents can query it instead of re-reading raw files.
Parameter descriptions lack format constraints and ranges. E.g., 'recent' in agentsmcp_receive states 'default 3, max 50' in description but no min/max schema constraints visible. 'limit' parameters similarly underconstrained.
Tool composition gaps. agentsmcp_send and agentsmcp_reply_all both send messages but differ in scope (single recipient vs. all participants). Descriptions do not clearly explain when to use each, risking LLM confusion.