AI access

Let agents help with credential handoffs without dumping secrets.

Agent-facing CLI and MCP commands list safe metadata, request an operation, and poll its status. They never accept or return passwords, TOTP setup links, seeds, codes, encrypted payloads, wrapped keys, or approval decisions.

Create workspace View plans

Agent session

$ secretseal vault list --workspace ops

18 metadata records returned

$ secretseal totp approval-request --id shared2fa-01

request created for human review

$ secretseal totp approval-status --workspace ops

waiting device approval

CLI for operators

List credential metadata, inspect Secure Send / Receive queues, and request Shared2FA approval from a local terminal.

Workspace-scoped API keys

Metadata-first commands

Metadata-only attention queues

Shared2FA approval requests

MCP for AI agents

Give Codex, Claude Code, and trusted local agents a controlled interface for credential handoffs under the same policy model humans use.

Tool-scoped permissions

No secret-shaped arguments

Human-only approval decisions

Shared 2FA guardrails

Agents can list safe account metadata, request human approval, and poll safe status without receiving setup material or codes.

Approval requests

Safe status polling

Human-only decisions

AI product loop

Agents help across the product without receiving secrets.

The same MCP CLI surface supports safe vault metadata, Secure Send / Receive queue planning, Shared2FA approval requests and status polling, and launch proof checks.

Vault metadata

List items, collections, owners, review dates, and rotation state while encrypted payloads stay sealed.

Secure exchange operations

Summarize active, expired, exhausted, and guarded sends plus open and completed receive requests without raw files or private notes.

Shared2FA approvals

Inspect authenticator records, request approval, and review audit state without exposing seeds or generated codes.

Launch proof handoff

Read paid billing, native parity, and live-agent proof gates before making sales or completion claims.

Permission ladder

Agents start useful, then ask before sensitive work.

SecretSeal separates agent metadata and request workflows from human-only decisions and secret-handling web workflows.

Safe now

13 actions

Agents can bootstrap, inspect readiness, list safe metadata, plan metadata-only work, request Shared2FA approval, poll status, and read launch proof handoff.

secretseal_agent_bootstrap

secretseal_status

secretseal_doctor

secretseal_key_scopes

secretseal_vault_list

Request and poll

Human decides

Agents may request Shared2FA approval and poll safe status metadata; approval decisions remain human-only.

secretseal_totp_approval_request

secretseal_totp_approval_status

Approval decisions remain human-only and are unavailable to CLI/MCP agents.

Unavailable to agents

No secret tools

Passwords, setup links, seeds, codes, encrypted payloads, wrapped keys, reveal/export paths, and approval decisions are absent from CLI/MCP.

Unknown names fail before client helpers run

Undeclared arguments fail before network requests

Connect an AI agent

Six setup steps, no token printed.

Operators can connect Codex, Claude Code, or local automation with scoped workspace keys while keeping secret-bearing paths unavailable to the agent surface.

Create a scoped workspace key

Open Settings -> API keys -> Metadata agent

Keep the token in an MCP client secret manager, OS keychain, or SECRETSEAL_API_KEY environment variable; record its non-secret source as keychain:<name>, secret-manager:<name>, ci-secret:<name>, or mcp-secret:<name> and attach the runtime handle only after a real production run.

Generate MCP client template

secretseal mcp client-template --client codex

The template names commands, required env vars, and the safe-first tool surface without printing token values.

Bootstrap before work starts

secretseal mcp bootstrap --client codex

Agents see readiness, the six approved scopes, two presets, unavailable boundaries, and safe approval request/status paths before work starts.

Run live agent proof drill

npm run agent:proof-template && npm run agent:launch-proof

Print safe env template, verify live API URL, live wrk_... workspace id, production environment, key presence, package entrypoints, and no-secret response boundaries.

Review exchange attention queues

secretseal exchange ops-plan --workspace <workspaceId>

Give agents Secure Send / Receive counts and recommended next actions without plaintext, ciphertext, wrapped keys, raw files, submitted files, or destructive actions.

Hand off product proof

secretseal launch proof-handoff

Give agents a metadata-only checklist for paid SaaS billing, native Authenticator parity, and live MCP/CLI proof before acting.

Live proof handoff

Promote MCP/CLI only after a production runtime proof handle exists.

Local metadata checks are available now. Live AI-agent access stays partial until the configured MCP or CLI runtime returns a non-secret production proof handle.

Accepted handles

agent-runtime:<host>:workspace-<hash>:run-<opaque-production-receipt>

SECRETSEAL_API_KEY_PROOF_SOURCE=keychain:<name> or secret-manager:<name>

SECRETSEAL_API_KEY_PROOF_SOURCE=mcp-secret:<name> or ci-secret:<name>

SECRETSEAL_AGENT_RUNTIME_PROOF=agent-runtime:<host>:workspace-<hash>:run-<opaque-production-receipt>

Current live-agent proof snapshot

live-mcp-cli-agent-access3/5 ready

Set a public production HTTPS API URL, secret-managed API key proof, live wrk_... workspace id, and SECRETSEAL_ENVIRONMENT=production before using live agent access.

SECRETSEAL_API_URLSECRETSEAL_API_KEYSECRETSEAL_WORKSPACE_IDSECRETSEAL_ENVIRONMENTSECRETSEAL_AGENT_RUNTIME_PROOF

Run live-agent proof drill

Completion rules

Run npm run agent:launch-proof only after SECRETSEAL_API_URL, SECRETSEAL_WORKSPACE_ID, SECRETSEAL_ENVIRONMENT, and protected SECRETSEAL_API_KEY proof are configured.

Use only the opaque receipt issued by the successful production safe metadata-only run; the readiness report validates and echoes the supplied agent-runtime handle but never generates it.

Keep the live MCP/CLI claim partial until the proof handle points to a real production host, approved workspace, production environment, and secret-managed API key proof.

Agent-safe workflow

Useful AI work starts metadata-first.

SecretSeal gives AI enough safe context to help operators while leaving secret-bearing and human-decision actions outside the CLI/MCP catalog.

Bootstrap safely

Point Codex, Claude Code, or local automation at scoped workspace metadata before any secret path is available.

Inspect metadata

Let agents list vault items, Shared2FA records, rotation dates, and readiness state without returning payloads or codes.

Plan the work

Use metadata-only ops-plan tools for rotation, review, Secure Send / Receive attention queues, and mobile-parity planning.

Request and poll

Request Shared2FA approval, then poll safe status while the human decision remains outside the CLI/MCP surface.

Safe by default

Agent access follows same policy surface.

The selling promise is strongest when the CLI, MCP server, web app, and future native apps share one audited access model.

API keys are scoped by workspace and tool capability.

Agents can discover safe metadata without a reveal or export path.

Secure Send / Receive tools return queue metadata and recommended next actions only.

Shared2FA agents can request approval and poll safe status; people make approval decisions.

Secret-shaped arguments are rejected before client helpers or network requests run.

Agent readiness

Verify AI access contract before handing agents a key.

The public readiness endpoint proves MCP and CLI setup shape without exposing API keys, vault payloads, TOTP seeds, generated codes, or customer data.

Production exposes /api/v1/mcp/readiness with setup metadata only.

secretseal_agent_bootstrap is the safe first MCP action for Codex and Claude Code.

secretseal_exchange_ops_plan summarizes Secure Send / Receive attention queues without accepting payload-shaped plans.

secretseal_launch_proof_handoff returns launch proof tracks without secrets.

Reveal, export, generated-code, payload-shaped planning, and approval-decision tools are unavailable to agents.

Workspace keys

Create scoped keys in Settings.

Keep agent credentials separate from human sign-in and rotate with the same operational discipline as client secrets.

Open settings