In-process, OTP-supervised security runtime for MCP and agent-tool boundaries with deterministic policy, Agent Trust attestations, and signed evidence.
SigilGuard is a specialized security library for MCP/agent-tool boundaries, not a traditional MCP server. All 8 tools have descriptions (194 - 250 chars, within baseline 34 - 392 range) and are READ_ONLY, reducing risk. Input schemas are present but use Elixir type annotations (e.g., 'term', 'Context.t() | map() | keyword()') rather than JSON Schema, making them opaque to JSON-based MCP clients. Tool names are action-verb-based (guard_*, verify_*, attest_*, issue_*, confirm_*) and clearly scoped. However, parameter descriptions lack detail on constraints, enums, and valid value ranges. Output schemas are not documented. Error handling and recovery guidance are absent from descriptions.
Build and sign a tool_request attestation envelope.
Build and sign a tool_result attestation envelope.
Guard a tool request using manifest checks before runtime policy evaluation. Supported options: :manifests (map of tool name to pinned manifest), :require_manifest (block unknown tools), :allow_suspicious_params (boundary-policy escape hatch), :attestation (:off, :optional, or :required), :confirmation (:honor or :off), :confirmation_token, :confirmation_key, :consume_confirmation, :now.
Guard a tool result before returning output to a model. Enforces manifests, attestations, and confirmations around tool result boundaries.
Issue a confirmation token for a confirm-required tool request or result. Pass direction: :request (default) for tool calls and direction: :result for results.
Verify one resource entry returned inside an MCP resources/read result. Checks ui:// identifier, MCP Apps MIME type, text-or-Base64 content shape, pinned SHA-256 digest, bounded content size, and resource's declared CSP origins, browser permissions, and dedicated app domain.
Input schemas use Elixir type annotations (e.g., 'term', 'Context.t() | map()') instead of JSON Schema. JSON-based MCP clients cannot parse or validate these types. Schemas must be converted to proper JSON Schema with explicit type, properties, required, and description fields.
Parameter descriptions lack constraint details. E.g., 'opts' is described as 'Options for manifest checks...' but does not specify valid keys, value types, ranges, or enums. LLMs cannot infer valid option combinations.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | <=2025-11-25 | v2 |
Verify a refreshed tools/list result after the host receives a list-change notification. Every observed entry is revalidated against its pinned manifest.
Verify an observed manifest against the configured manifest set. Accepts a tool name string or observed manifest map.
Output schemas are not documented. Tool descriptions do not specify what fields are returned, their types, or structure. Agents cannot plan downstream calls or extract required data.
No error handling or recovery guidance in tool descriptions. Tools do not document what errors can occur, how to classify them (retryable vs. fatal), or what the LLM should do next.
Transport mechanism is unknown. Source code shows Elixir/OTP library structure but does not reveal how tools are exposed to MCP clients (STDIO, HTTP, SSE, or embedded). Cannot assess protocol readiness without transport clarity.