Agent runtime with full system prompt control. Define agents in markdown, run them in a terminal TUI or multi-user web UI. Powered by the Claude Agent SDK.
17 tools with basic schemas and descriptions, but significant gaps in parameter documentation and error handling. Most tools have 1-2 parameters with minimal descriptions (10-30 chars). No output schemas documented. No error recovery guidance. Naming is verb-noun compliant (memory_read, web_search, shell_exec) but descriptions lack WHEN/WHY context. High-risk tools (shell_exec, write_file) lack confirmation patterns. No tool annotations (readOnlyHint/destructiveHint). Composition is reasonable (separate read/write tools) but parameter descriptions are too terse to guide LLM selection reliably.
Call another agent via A2A protocol
Read the agent's identity document
List agent proposals
List files in a directory
Read from agent memory files
Write to agent memory files
List available Claude models
Read file contents from the workspace
Read scratchpad content
Parameter descriptions are too terse (10-30 chars). LLMs cannot infer meaning from bare names like 'filename', 'path', 'status', 'agentId'. Descriptions must explain format, constraints, and when to use each parameter.
No output schemas documented. LLMs cannot plan downstream tool calls or extract required fields (e.g., does task_list return task_id, title, status?). Every tool must document its return structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Write scratchpad content
Execute shell commands
Create a new task in the task list
List all tasks
Update a task status
Fetch and parse web content from a URL
Search the web for information
Write file contents to the workspace
Destructive tools (shell_exec, write_file, memory_write, task_create, a2a_call_agent) lack confirmation patterns or dry-run support. Agents can accidentally delete files, execute harmful commands, or corrupt state without safeguards.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). MCP clients cannot distinguish safe tools from dangerous ones. shell_exec and write_file should be marked destructive; memory_read should be marked read-only.
No error recovery guidance. Tools return errors but do not tell LLMs what to do next (retry, ask user, try alternative tool). E.g., shell_exec failure should suggest 'Check command syntax or permissions' or 'Try a simpler command first'.