Skip to main content
GhostStack
Menu
vault Status: Coming soon
Building

A credential vault for agents and humans alike.

Vault serves both populations from one store. Humans manage credentials through a clean UI. Agents request them through MCP with scoped, time-bound leases. Credentials live in sandboxed tool execution, not in LLM context windows, and every access is attributed to a specific agent run or human session.

Waitlist opens with launch. Check back when the page flips to Available.

01 / Credentials

What Vault holds

01

API keys

Plain string credentials with per-agent scope. Rotation windows, last-used timestamps, revocation one action deep.

02

OAuth tokens

OAuth 2.1 access + refresh pairs with auto-refresh. Provider catalog covers the integrations the suite uses; external providers are config.

03

SSH keypairs

Public/private pairs stored encrypted. Ed25519 by default; RSA when a host still requires it.

04

TLS certificates

Cert + key + chain together. Expiry surfaced on the dashboard; no surprises during a deploy.

05

Database URLs

Connection strings with driver awareness. Treated as single credentials so the whole string rotates together.

06

Generic secrets

Anything that is a string and should never sit in an env file. Same encryption envelope, same audit discipline.

02 / For both

Built for agents and humans alike

Password managers built only for humans do not answer the questions an AI agent asks. Vaults built only for apps do not give humans a dashboard. Vault serves both. The agent side is scoped leases over MCP. The human side is a clean UI for the same credentials.

01

MCP server, any MCP client

Vault exposes an MCP server following the 2025-06-18 spec. Claude Code, Cursor, Windsurf, or any Cortex agent can list scoped credentials and request them with a signed justification.

02

Credential leasing with TTLs

Requests return leases bound to the agent run id. Default TTL is 5 to 10 minutes, configurable per credential. Lease expiry revokes access without requiring the agent to clean up.

03

Agent attribution in audit log

Every access records which agent type, which user, which run, what justification. The audit log is append-only and survives tenant deletion so incident response has history to read.

04

Sandboxed delivery, not context injection

Credentials are delivered into a sandboxed tool execution environment. They never enter the LLM's context window. Prompt-injection cannot exfiltrate what the prompt never saw.

05

Per-agent-type policies

Scope credentials to specific agent types, not just users. A scoring agent that never needs production DB creds never gets asked about them.

03 / Browser

On every browser tab

A Chrome and Firefox extension brings Vault into the place humans spend most of their day: browser tabs that ask for a username and a password. Same encryption, same audit chain, same RBAC. No long-lived plaintext cache; each fill is a fresh server-side lease consumed in one round-trip.

01

Chrome + Firefox browser extension

A Manifest V3 extension for both browsers auto-fills usernames and passwords on the sites you log in to every day. Sign in once via identity OAuth in a popup; the extension reads stored `login` credentials over the same REST API your CLI uses.

02

Phishing-resistant URL matching

Match is scheme-strict (HTTPS only by default), Punycode-normalized to defeat homograph attacks, and port-aware. Subdomain match is opt-in per credential. The fill button only appears when the URL passes every check.

03

Lease-per-fill, no plaintext cache

Every fill issues a fresh 30-second server-side lease, single-use. The plaintext lives in the extension service worker for one round-trip and is overwritten immediately after the form is populated. No encrypted-on-disk cache.

04

Same audit trail as every other access

Each fill writes lease.issued, lease.consumed, lease.read to the tamper-evident audit chain with the user attribution and a session correlation id. Indistinguishable from a CLI reveal or a manual web-ui copy.

04 / Comparison

How Vault compares

Capability matrix
05 columns / 06 rows
capability Vault 1Password Bitwarden Doppler HashiCorp Vault
MCP server for agent access Yes No No No No
Credential leasing with TTL Yes No No No Yes
Per-agent-type policy scope Yes No No No No
Personal + team credentials Yes Yes Yes No No
Runtime secret injection Yes No No Yes Yes
Target user Developers with AI agents Humans Humans Apps at runtime Enterprise infra

05 / Security

Security posture

  1. Claim 01
    of 06

    AES-256-GCM envelope encryption

    Each credential is sealed with a per-credential data encryption key (DEK). DEKs are wrapped by a per-tenant key encryption key (KEK). Compromise of storage media alone is not enough to read a credential.

  2. Claim 02
    of 06

    Pluggable KEK (KMS or customer HSM)

    KEK lives in AWS KMS or GCP KMS by default. Enterprise tenants can bring a customer-managed HSM. Rotation is a KMS operation; envelope decryption continues through a rotation without downtime.

  3. Claim 03
    of 06

    Append-only audit log

    Every access, lease, grant, and revoke event writes to an audit log that outlasts tenant deletion. Operators and compliance reviewers read the same log; there is no cleaner view that redacts entries.

  4. Claim 04
    of 06

    Tamper-evident integrity

    Stored ciphertexts include an integrity tag verified on every read. Silent corruption is detected rather than returned. A read that fails integrity fails loud.

  5. Claim 05
    of 06

    Tenant isolation at three layers

    Postgres row-level security policies, branded TenantId types at the application layer, and a tenant-scoped repository decorator. Defense in depth, not a single choke point.

  6. Claim 06
    of 06

    OAuth 2.1 with Identity

    Every Vault session is an OAuth 2.1 session issued by Identity. Cross-sibling single sign-on with rotating JWTs and a refresh mechanism.

Also in the suite