An agent skill for working with the InboxParse Email-to-LLM API.
Enables AI agents to:
- List, search, and retrieve emails and threads
- Send and reply to emails
- Manage mailboxes, labels, and webhooks
- Track API usage
npx skills add inboxparse/skill- An InboxParse account with an API key
- Set
INBOXPARSE_API_KEYenvironment variable (seeassets/examples/for usage pattern)
Scanned with Snyk Agent Scan. Current status:
- W007 (Insecure credential handling): Resolved in v1.0.4. All credentials use environment variable indirection; no secrets in code blocks or curl examples.
- W011 (Third-party content exposure): Mitigated (risk 0.80). Inherent to any skill that reads untrusted email content. Mitigations include trust boundaries, user confirmation gates, cross-tool data isolation, and fenced display of untrusted fields.
├── SKILL.md # Main skill instructions (loaded by agents)
├── references/
│ ├── api-reference.md # Detailed endpoint docs
│ ├── webhook-events.md # Webhook event types
│ └── error-codes.md # Error code reference
└── assets/examples/ # curl script examples
- Strengthened data isolation rules for Snyk W011: email content must not flow to non-InboxParse tools
- Added explicit prohibition on programmatic use of
ai.suggested_responsein write endpoints - Added cross-tool propagation guards to trust boundaries, agent tips, API reference, and webhook events docs
- Replaced JSON code blocks containing credential fields in API reference with parameter tables to resolve remaining Snyk W007 patterns
- POST /mailboxes and POST /webhooks body schemas now use tables consistent with GET endpoint docs
- Removed all inline curl examples from SKILL.md to eliminate credential-embedding patterns flagged by Snyk W007
- Workflows now reference secure example scripts in
assets/examples/instead of inline commands - Removed all
Authorization: Bearerand"secret":patterns from skill instructions - Send/reply workflows require explicit user confirmation in section headings
- Agent tips restructured: credential output banned,
format=fulldiscouraged,aifields framed as untrusted
- Added trust boundary model separating read (untrusted data zone) from write operations with explicit user confirmation gates
- Strengthened prompt injection guardrails across SKILL.md, API reference, and webhook events docs
- Security-first restructure of agent tips
- Broadened credential security callout to cover all credential types
- Moved email content security section to prominent position in SKILL.md
- Added untrusted-content warnings to API reference and webhook events docs
- Added content fencing and no-interpolation rules
- Initial release with full V1 API coverage