What was removed, and why

The rebuilt control plane is not a refactor of the proof-of-concept. It is a ground-up rebuild around a much narrower scope, and it deleted rather than reimplemented a large part of the previous product.

Nothing needed a deprecation period: no production traffic depended on any of it.

The principle

Do less, better. VaultysClaw's job is agent identity and trust. Everything below was either better served by a tool built for it, or existed only to support something else on this list.

A secondary principle, applied throughout: delete code rather than patch it. Several concrete security holes — a stubbed signature verifier that always returned true, plaintext bridge secrets — were closed by removing the feature that contained them, not by fixing it.

What went

Workflow engine

Executor, scheduler, approvals, runs, steps. Deferred to n8n or similar.

Orchestration is a solved problem with mature tools. Re-implementing a smaller, worse version of one was not a good use of the platform's scope, and the orchestrator does not need to be the same system as the trust plane — an n8n instance can be governed as an Actor like anything else.

Channels, chat, threads, mentions

The entire collaboration surface. It existed so humans could converse with agents conversationally, and it went with the end-user product.

Partially replaced by the Access Portal: a certificate viewer plus a launch point. Any human holding a connect grant can reach the agent that grant names — what "connecting" means is owned by that agent's kind. That is narrower than the channel system on purpose, and it is not a staging post back toward it.

Teams and generic bridges

Existed only to sync the channel system externally. Removed with it — which also removed a stubbed request verifier that always returned true and a set of plaintext-stored bridge secrets.

In-app, email, and push notifications

The whole notifier package, notification rows, and per-user preferences. There is no general end-user population to show a bell to.

Replaced by notification channels: operator-facing alerts through a self-hosted Apprise container, driven by the same event pipeline as webhooks.

The end-user settings area

Profile, security, appearance, notification preferences. Partially replaced by a first-login profile prompt and the Access Portal.

The ts-rest REST API

Every REST route, the contract layer, and the hashed bearer-token API keys.

There is currently no REST API surface at all. The admin console is Server Actions over data-access objects. The replacement design is a service DID holding a scoped certificate and signing each request — not a bearer token that works for whoever holds it.

Why the old API keys were not ported

Beyond the design difference: the previous package's API-key administration routes never called the authorisation helper at all, and were therefore unauthenticated. Porting them would have carried that across. Whoever builds the replacement should start from the design, not from that code.

Three parallel grant models

User grants, delegation certificates, and agent peer grants — three models each implementing a variation of "grant a subset of capabilities to someone, signed, with expiry", each with its own revocation and expiry logic.

Collapsed into one thing: a CapabilityCertificate row. A human delegating scoped access to another human, or an agent granting a peer access, is the same artefact — the subject is any Actor's DID, and the scope expresses the narrowing. One ledger, one status protocol, one audit trail, instead of four models each needing their own.

The role enum

Owner, Admin, Member. Replaced by capabilities in the same ledger: admin_console_access, portal_access. One fewer parallel system, and one more thing auditable through the single ledger rather than a role table off to the side.

Two separate device models

Agent and SensorDevice were two Prisma models with two admin pages and two structurally identical registration flows. Collapsed into Actor, with kind-specific configuration in one JSON column.

Two separate audit logs

An intent log and an activity log. Merged into one, since there is no separate "intent execution" concept to track.

What was explicitly kept

• The VaultysId crypto layer — passwordless QR login, the Challenger handshake, the certificate primitives. This was never the problem.
• Workspaces — still the tenancy boundary, scoped now to certificates, budgets, and model access rather than also workflows and channels.
• OIDC and Entra ID — as identity-establishment paths that bind to a DID, never as a parallel trust tier.
• The model registry and LiteLLM routing.
• The credentials vault — extended to cover Apprise service URLs, replacing the inconsistent plaintext path the bridges used.
• Webhooks — kept as core, and extended to carry what used to be notifications.

Connecting was never in scope for removal

Every reduction is about what the interface does once you are connected, never how you connect.

The passwordless VaultysId QR login stays exactly as it was. No username-and-password fallback was introduced just because the audience narrowed to operators — the point of the rebuild is fewer surfaces with more trust rigour, not less. The agent registration and certificate handshake is likewise untouched: the Actor model changes what happens after a successful handshake, never the handshake itself.

Coexistence

The rebuilt control plane lives alongside the older one in the same repository. Nothing points production traffic at the rebuild until a cutover is deliberate. They use different ports and different databases, and can run side by side.

Docs on this site describe the rebuilt control plane. Where the older package's documentation disagrees, this site wins.