Secret management for AI coding agents. Your secrets never enter the LLM context window.
VaultBridge demonstrates strong security-first design with explicit tool annotations and clear descriptions. All 4 tools have descriptions (194-250 chars, within baseline 34-392 range) and input schemas with type definitions. Tool names follow verb_noun pattern (vault_search, vault_status, vault_store, vault_inject). However, output schemas are not formally documented, responses are inferred from code (JSON text content). Error handling is present but inconsistent: vault_store returns structured error JSON for non-generate case, but other tools lack recovery guidance. Parameter descriptions are good (50-80 chars) but some lack format constraints (e.g., expires_at accepts ISO 8601 but no pattern validation shown). Security is excellent: no secrets in parameters, server-side injection enforced. Composition is clean: each tool has one responsibility, and vault_inject properly chains with vault_search/vault_store.
Inject a secret value into a target without the value entering the LLM context. Supports: file (.env/yaml/json), clipboard (auto-clear), template ({{VAULT:ref}} placeholders). For remote callers, returns a one-time redeem token instead of injecting directly.
Search the vault by name, service, project, environment, type, or env var name. Returns metadata arrays ONLY — never secret values. Use vault_inject to access actual values.
Check if the vault is unlocked and accessible. Returns connection status.
Store a new secret in the vault. The actual secret value must be provided via a secure channel (auto-captured by hook, clipboard, or stdin). If generate=true, a random password is generated and stored. Returns confirmation with ID — never the actual value.
Output schemas not formally documented. Tool responses return JSON text content, but the structure of that JSON (fields, types, required vs optional) is not declared in tool definitions. LLMs cannot plan downstream calls or extract fields reliably without knowing response structure.
vault_status has minimal description (75 chars). 'Check if the vault is unlocked and accessible. Returns connection status.' lacks context on when to call it vs vault_search, what 'accessible' means, or what fields the response contains. Should explain: 'Call before any vault operation to verify the vault is unlocked. Returns locked status, authenticated email, and server URL. If locked, user must run rbw unlock or bw unlock.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
vault_inject 'local' parameter lacks clear guidance. Description says 'True if caller is on the same machine as the broker' but does not explain the consequence: when local=true, secret is injected directly; when false, a redeem token is returned. This dependency should be explicit: 'Set to true only if the caller is on the same machine as the vault broker. When true, secret is injected directly to the target. When false, a one-time redeem token is returned instead (for remote callers).'
vault_store non-generate path returns error JSON but does not guide recovery. Error message says 'Cannot accept secret values directly via tool parameters' but does not suggest alternatives (e.g., 'Use the auto-capture hook by pasting the secret into a terminal command, or call vault_inject with an existing vault item'). Error responses should include actionable next steps.
vault_search result limit not documented. Response includes 'total' count but no indication of max results returned or pagination support. If vault contains 1000 secrets, does search return all 1000 or cap at 50? Baseline pattern requires pagination or explicit limit statement in description.