MCP server for remote command execution, file operations, and job management on distributed agents via S3-compatible object storage
Strong foundation with 23 well-named tools following verb_noun convention (list_agents, exec, read_file, etc.). All tools have descriptions (avg ~120 chars, within baseline 34-392 range). Input schemas present for all tools with typed parameters and descriptions. However, output schemas are entirely undocumented, no tool specifies what fields it returns, forcing LLMs to infer structure. Error handling is minimal; most tools lack recovery guidance. Security-sensitive tools (exec, write_file, remove) lack explicit permission/audit documentation. Some parameter descriptions are terse (e.g., 'Agent identifier' repeated 20+ times without context on how to obtain agent_id). No tool uses enums for constrained inputs (e.g., binary flag in read_file could be boolean enum). Composition is sound, tools chain well (list_agents → ping → exec), but no batch variants for common loops.
Cancel a running job.
Run one program on the agent. No shell, arguments never expanded. An agent runs commands one at a time, so a slow one delays everything queued behind it — use start_job for anything past a few minutes.
Get recent output from a job's stdout and stderr streams.
Get status of a specific job.
List live agents: version, running jobs, recent errors, CPU and GPU metrics.
List directory contents on the agent.
No output schemas documented for any tool. LLMs cannot infer what fields list_agents, job_status, or pull_file return, forcing them to guess structure and risking failed downstream chaining.
Destructive tools (remove, share_remove, stand_down) lack confirmation/dry-run patterns and do not document permission requirements or audit trails. No guidance on recovery if called in error.
Parameter descriptions are generic and repetitive ('Agent identifier' appears 20+ times without explaining how to obtain agent_id or what happens if it is invalid). No actionable error guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2026-07-28+ | v2 |
List all jobs on an agent, running and recently finished.
Create a directory on the agent.
Move or rename a file or directory on the agent.
Round-trip check that one agent is alive.
Publish a new relay-agent binary for agents to download and install.
Download a file from an agent to the controller machine.
Upload a file from the controller machine to an agent.
Read a file from the agent. Text validates UTF-8, binary returns base64. Above ~100 KB use pull_file instead: this lands in the conversation.
Delete a file or directory on the agent.
Withdraw a published agent binary update.
Retrieve a file from the shared area to an agent.
List files available in the shared area.
Make a file from one agent available to others via a shared area.
Delete a file from the shared area.
Stop a specific agent instance from accepting commands.
Start a long-running job on the agent that runs detached and reports completion via heartbeat.
Write a file on the agent.
No enum constraints on boolean or choice parameters (e.g., binary flag in read_file, append flag in write_file, recursive flag in remove). LLMs may pass invalid values.
Security-sensitive tools (exec, write_file, remove, publish_update) do not declare required permissions or scope (e.g., 'exec:untrusted', 'write:filesystem'). No audit trail or permission-gate documentation.