Loading ToolBench
MCP server implementation for AgentCell platform operations, providing tools for deploying cells, managing logs, rollbacks, environment variables, secrets, domains, access control, and resource monitoring
11 tools with complete schemas and descriptions, but significant quality gaps. All tools have input schemas with typed properties and descriptions (good baseline). However, descriptions are terse (avg ~40 chars, well below 194-char baseline), lack actionable context, and omit prerequisites/side effects. Parameter descriptions are minimal. No enum constraints on action parameters (env, secrets, domains, access accept free-form 'list'/'set'/'unset' strings). Error handling guidance absent. Destructive operations (destroy, rollback, deploy) lack confirmation patterns or dry-run support. Tool naming is verb-first (good), but some names are vague: 'env' and 'secrets' don't clarify they modify state; 'spend' doesn't indicate it's read-only. No output schemas documented. Composition is reasonable (single responsibility per tool), but critical chaining IDs missing from responses (e.g., logs tool doesn't document what fields it returns).
List or revoke access grants
Archive a source directory and deploy it as a cell
Destroy a cell while retaining recoverable backups. Destructive: requires a separately scoped token; confirmation is enforced where present.
List or change custom domains
List or change non-secret environment variables
Stream build or runtime logs for a cell. Returns a stream.
List running and sleeping cells
Action parameters (env, secrets, domains, access) accept free-form strings instead of enums. LLMs will hallucinate invalid actions like 'modify', 'delete', 'get' instead of the required 'list'/'set'/'unset'.
Descriptions are terse (avg 35 chars, baseline 194 chars) and lack actionable context. 'List or change environment variables' doesn't explain when to use env vs secrets, what happens on set/unset, or prerequisites. LLMs cannot reliably select the right tool.
Destructive operations (destroy, rollback, deploy) lack dry-run or confirmation patterns. destroy requires a 'confirm' parameter, but no guidance on what happens if confirmation fails or how to preview changes. Agents risk irreversible data loss.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Restore a previous successful deployment
List or change secret names without returning secret values
Grant a user or group access to a cell
Report consumption spend
No output schemas documented. logs returns a stream, but the response structure is not specified. deploy, rollback, and others don't document what fields they return, forcing LLMs to guess what data is available for downstream tool calls.
Tool names 'env' and 'secrets' are ambiguous, they don't signal that these tools modify state. 'set_env' and 'set_secrets' would be clearer. 'ps' is jargon (Unix process status); 'list_cells' would be more discoverable.