Loading ToolBench
Three-tier LLM agent orchestration with persisted types, earned trust, and symmetric supervision. Exposes tools for managing projects, runs, skills, registry, and platform state over MCP HTTP transport.
The Atoma MCP server exposes 23 tools with basic descriptions but critical gaps in schema documentation and parameter clarity. Tool names follow verb_noun convention (atoma_run_start, atoma_skills_list) which is good, but descriptions are inconsistent in depth. Most critically: source code provided does not show explicit input schemas, parameter types, or output schema documentation for any tool. Without visible schema definitions, parameter descriptions, or structured output specs, the server cannot reliably guide LLM tool selection or parameter binding. Descriptions range from adequate (e.g., 'Start a new run with optional task augmentation') to vague (e.g., 'Show details of a verdict'). No evidence of error handling guidance, recovery paths, or actionable error messages. Risk classifications (READ_ONLY, WRITE, DESTRUCTIVE) are present but not integrated into tool annotations.
Show cost information for runs and operations
List task families and their goal templates for phrasing goals
Get the friction report showing operational issues
Check the integrity of the platform ledger
Get recent entries from the platform ledger
Show the history of an agent type's trust and promotion
No input schemas visible in source code. All 23 tools lack documented parameter types, constraints, and validation rules. LLMs cannot infer valid parameter values or ranges.
Descriptions are generic and lack actionable context. Examples: 'Show details of a verdict', 'List verdicts on the platform' do not explain WHEN to call the tool, what it returns, or how it differs from related tools.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2025-06-18+ | v2 |
List agent types in the platform registry
Rollback an agent type to a previous version (platform only)
Show details of an agent type in the registry
Cancel an active run
Start a new run with optional task augmentation; returns immediately or when run ends depending on augmentation
Get the trace of a run showing agent decisions and tool invocations
Get a specific trace file from a run
List runs in the organisation or platform
Remove a skill from the platform catalog (platform only)
Merge a skill into the platform catalog (platform only)
Reset a skill's trust and promotion state (platform only)
List skills in the platform skill catalog
Mechanical pre-screen of a skill (not a sharing approval)
Show details of a skill including its body and trust status
Statistics on skills including trust and promotion status
List verdicts (skill promotion decisions) on the platform
Show details of a verdict
No output schema documentation. Tools return data but LLMs cannot plan downstream calls without knowing response structure, field names, or available IDs for chaining.
Risk classifications (READ_ONLY, WRITE, DESTRUCTIVE) are metadata but not integrated as tool annotations. MCP spec supports readOnlyHint, destructiveHint, idempotentHint, these should be present in tool definitions.
No error handling guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next (e.g., 'Run not found, try atoma_runs_list() first').