Loading ToolBench
The file protocol for AI agent context. Encrypted, signed, peer-to-peer. Provides HTTP daemon for serving context store over A2A (agent-to-agent) messaging and peer synchronization.
OpenFused MCP exhibits significant definition quality gaps. While tool names follow verb_noun conventions (send_message, list_tasks, get_task), descriptions are inconsistent, some are adequate (50-100 chars) but many lack depth. Critical issue: input schemas are visible but parameter descriptions are sparse or missing entirely. For example, 'send_message' accepts a 'message' object but provides no detail on required fields (from, to, content, encryption). Output schemas are undocumented, LLMs cannot predict what fields to expect from responses. Error handling is absent from descriptions. The server implements 17 tools but lacks the structured, LLM-optimized documentation needed for reliable agent composition. Naming is generally clear, but schema completeness and parameter guidance are weak.
Acknowledge (delete) an outbox message after successful delivery
Cancel a running task by ID
Create an artifact (output file) for a task
Retrieve the agent's discovery card (name, description, capabilities, skills)
Retrieve the context store configuration (ID, name, public key, encryption key)
Retrieve outgoing messages from the agent's outbox (for peer sync)
Retrieve the agent's PROFILE.md (public address card in markdown)
Input schemas lack parameter descriptions. Tools like 'send_message' accept complex objects (message envelope with from, to, content, encryption) but provide no guidance on required fields, formats, or constraints. LLMs cannot infer what to pass.
Output schemas are completely undocumented. No tool description specifies what fields the response contains, their types, or how to chain results to downstream tools. Agents cannot plan multi-step sequences.
No error handling guidance. Tool descriptions do not explain what errors can occur, whether they are retryable, or what the LLM should do next. A failed 'send_message' call provides no recovery path.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Retrieve a specific task by ID with full details including input, status, and events
List files in a directory (shared/, knowledge/, etc.) — full mode only
List root-level files and directories (CONTEXT.md, PROFILE.md, shared/, knowledge/) — full mode only
List all tasks in the context store, sorted by most recently updated first
Read a file from the context store (CONTEXT.md, PROFILE.md, shared/, knowledge/) — full mode only
Receive an incoming message in the agent's inbox (OpenFused signed message format)
Send an agent-to-agent message to another agent's inbox
Stream agent-to-agent messages with server-sent events (SSE)
Subscribe to task status updates via server-sent events (SSE)
Update the status of a task (e.g., running, completed, failed)
Pagination and result limits are not documented. 'list_tasks' and 'list_dir' do not specify max results, pagination parameters, or whether results are truncated. Large result sets could exhaust context windows.
Tool descriptions are generic and lack WHEN/WHY guidance. 'Stream agent-to-agent messages with server-sent events' does not explain when to use stream_message vs send_message, or what the LLM should do with streamed results.