A deterministic, self-healing egress gate (MCP server) for locked-down LLM agents. Gives LLM agents internet access without letting them roam or exfiltrate by driving a real, logged-in Chrome over CDP with hard guarantees: single egress domain, fail-closed behavior, one-way data flow, and status-only audit logging.
Airlock defines 3 tools with clear, security-focused descriptions (50-150 chars each). All tools have input schemas with typed parameters and descriptions. However, output schemas are not formally documented, responses are inferred from code. Parameter descriptions lack constraint details (e.g., max query length, URL format validation rules). Error handling returns status codes but lacks recovery guidance. No tool annotations (readOnlyHint, idempotentHint) despite all tools being read-only. Naming is clear (gate_status, search, fetch_page) and follows verb_noun convention. Composition is sound: each tool has one responsibility, and outputs include request_id for audit chaining.
Fetch the rendered TEXT of a public page (read-only, one-way). Disabled unless AIRLOCK_ALLOW_FETCH=1. Fail-closed if the gate is down; only http(s) URLs.
Is the egress aperture (the dedicated Chrome) reachable? Auto-launches it if down. This is the G1 precondition every outbound call checks first.
Run ONE query against the single allowed egress domain and return RAW results one-way. Fail-closed: if the gate is down, NO outbound call is made.
Output schemas not formally documented. Responses are inferred from code (dict with request_id, status, ts, results/content/url). LLMs cannot plan downstream tool calls without knowing the exact response structure.
Parameter descriptions lack constraint details. 'query' has no max length; 'url' has no format validation rules stated; 'request_id' has no format (UUID? alphanumeric?). LLMs cannot validate inputs without explicit constraints.
Error responses lack recovery guidance. Status codes like 'gate_offline', 'provider_error', 'bad_request' are returned but do not tell the LLM what to do next (retry? call gate_status first? ask user?). Per pattern:recovery-guide, errors must guide the agent's next step.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
No tool annotations despite all tools being read-only and idempotent. readOnlyHint and idempotentHint are current MCP patterns (2026-07-28) that help agents reason about safety and retry logic.
search() and fetch_page() do not document pagination or result limits in parameter descriptions. MAX_RESULTS=10 and FETCH_MAX_CHARS=15000 are enforced in code but not exposed to the LLM via parameter constraints or descriptions.