Reliability + security proxy for the Model Context Protocol (MCP): self-healing connections, runtime tool-security (rug-pull/poisoning detection), and a compliance-mapped audit trail.
mcp-bastion exposes 4 control tools with strong, clear descriptions and well-defined input schemas. Tool naming follows verb_noun convention (bastion__status, bastion__reconnect, bastion__security, bastion__compliance), and descriptions are 150 - 250 characters, within the 10 - 1024 char baseline and detailed enough for LLM selection. All tools have explicit JSON Schema inputSchema definitions with type, properties, and additionalProperties constraints. However, the server lacks output schema documentation, no schema definition for what these tools return (what fields, structure, types). Tool descriptions state WHAT the tool does and WHEN to call it but do not explicitly document the return structure. Error handling guidance is absent: tools do not specify which errors are retryable, user-fixable, or fatal. The 'bastion__compliance' tool conditionally requires audit to be enabled but does not explain what happens if audit is disabled or how to enable it. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are present despite clear semantic intent (Status/Security/Compliance are read-only, Reconnect is write-like). Parameter descriptions in 'bastion__reconnect' are minimal ('Name of the upstream server to reconnect (as shown by bastion status)'), good context but could benefit from examples or error guidance ('If reconnect fails, check server logs or bastion__status for details'). Missing: sample error messages, output field documentation, security considerations (e.g., what if an agent calls bastion__reconnect in a loop?), and composition guidance (how these tools chain together).
Report an audit/compliance summary of recent tool activity, mapped to governance frameworks (NIST AI RMF functions and OWASP LLM Top 10 categories). Requires audit to be enabled.
Attempt to reconnect a disconnected or failed MCP server by name. Use this after bastion status shows a server is not connected, to restore its tools without human intervention.
Report the security state of every proxied tool: whether its definition matches what was first approved (or changed — a possible 'rug pull'), any poisoning heuristics its description triggered, and whether another server exposes a tool with the same name (shadowing).
Report the health of every MCP server proxied by Bastion: which are connected, disconnected, reconnecting, or failed; how many tools each exposes; and the last error. Call this when a tool seems to be missing or a server may have dropped mid-session.
No output schema documentation. Tools return data but callers (LLMs, clients) have no formal specification of response structure, field types, or required vs optional fields. This forces LLMs to guess at response shape and risks misinterpretation of compliance/security reports.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Status/Security/Compliance are semantically read-only; Reconnect has side effects. Annotations explicitly declare safety/mutation intent to agents, preventing accidental misuse.
Missing error handling guidance. No specification of error conditions, retryability, or recovery steps. E.g., what if bastion__reconnect fails? Is it transient (retry) or indicates misconfiguration (ask operator)? No error classification in tool descriptions.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
bastion__compliance description mentions 'Requires audit to be enabled' but does not specify what response to return if audit is disabled, how to enable it, or whether the agent should take corrective action.
No rate-limit or runaway-protection guidance. If an agent calls bastion__reconnect repeatedly in a loop, there is no documented behavior (backoff, rejection, throttling). This risks overwhelming Bastion's proxy layer.