Configurable bridge for external agents to delegate actions to Hermes Agent via CLI or MCP.
The Hermes Action Bridge defines 8 tools with clear semantic structure and descriptions, but has significant gaps in parameter documentation and output schema clarity. Tool descriptions are moderately detailed (90-150 chars), meeting the baseline for discovery, but parameter descriptions are sparse or missing, and output schemas are not documented. The naming convention is consistent (hermes_* prefix with action verbs: capabilities, plan, prepare, execute, draft, approve, reject, request_approval), which is good for discoverability. However, several parameters lack descriptions (e.g., 'contextFiles' array items, 'provider', 'model'), and no tool documents its return schema or pagination behavior. Error handling guidance is minimal, tools declare risk levels (READ_ONLY, WRITE, IRREVERSIBLE, REVERSIBLE) but do not explain recovery steps or actionable error messages. The approval workflow tools (hermes_approve, hermes_reject, hermes_request_approval) form a good composition pattern for managing side effects, but lack clear sequencing documentation. Overall: solid foundation, but underinvested in parameter detail and output documentation.
Approve a prepared action using its approval ID.
Retrieve capabilities and configuration of the Hermes Action Bridge, including runtime status, configured presets, and policy settings.
Draft an action with Hermes in draft mode, allowing iteration without external side effects.
Execute an action with Hermes. Requires interactive confirmation from the client for external actions.
Plan an action with Hermes without executing it. Returns the plan that would be executed.
Prepare an action for execution with Hermes, obtaining an approval ID. The action is not executed until approved.
Reject a prepared action using its approval ID.
Output schemas are not documented for any tool. LLMs cannot determine what fields to expect from responses, forcing them to infer structure or make unsafe assumptions about return values.
Multiple parameters lack descriptions: 'contextFiles', 'provider', 'model', 'maxTurns', 'timeoutSeconds' are either undescribed or have minimal guidance on format and constraints. This violates the pattern that every parameter must have a non-empty, contextual description.
No error recovery guidance in tool descriptions. Tools declare risk levels (IRREVERSIBLE, WRITE) but do not explain what errors are possible, which are retryable, or how to recover. Error handling must guide the LLM's next step.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 56 | <=2025-11-25 | v2 |
Request approval for an action with Hermes using request-approval mode.
hermes_execute accepts a 'yolo' boolean parameter to 'bypass bridge policy' but does not document the security implications, authorization requirements, or audit trail. This is a dangerous control that lacks safeguards.
No pagination support documented for tools that may return lists (hermes_capabilities, hermes_plan). If these tools return multiple items, LLMs need to know the limit, offset/cursor, and total count to avoid context window exhaustion.
Parameter constraints are not formalized. 'maxTurns' and 'timeoutSeconds' lack min/max bounds, allowing LLMs to pass invalid values. These should declare ranges (e.g., maxTurns 1-100, timeoutSeconds 1-3600).