Only documented APIs
Every endpoint is named in the connector’s page. Where a service offers no API, that is written down and the work stops. Nothing is scraped or guessed.
For developers
Extend it, without loosening it.
Connectors, skills, your own endpoints and the Model Context Protocol. Each one passes through the same tiers and the same approvals as the steward itself.
I.
Connectors
One contract for every service.
A connector is TypeScript written against one contract: how it signs in, what it reports, what it can do and at what tier, and how it knows it is healthy. Ours are built with the same kit, @castald/connectors-core, and held to the same checks.
A tier on every tool
Reads are T0. Anything that touches money, messages, access or deletion asks first, and says in plain words what it will do. The checker fails a connector that does otherwise.
Secrets through the vault
A connector reads its keys only through the context it is handed. No environment, no files, no literals, and every read is audited.
Every event, once
Each event carries a key of its own, so a poll that repeats itself never repeats on the timeline.
Tested without a network
Recorded responses stand in for the service. Tests run in a moment and never touch a real account.
Your account, your guide
Each connector carries a setup guide: which account the household needs, where to make it, what to paste and who pays. It is shown before Connect.
Sign-in kinds
OAuth 2
API key
Local token
Username and password
MCP
II.
Skills
Skills are signed data, never code.
A skill arranges what the steward can already do: tools composed from the household’s own, a few lines of guidance, routines and tiles. It never receives a secret or a raw event.
-
i.
Submitted
Checked on arrival: every step must name a tool that exists, and no tier may sit below its riskiest step.
-
ii.
Reviewed
Read by our staff, step by step, before anything is offered.
-
iii.
Signed
Published with an Ed25519 signature that is checked again before every install.
-
iv.
Installed
The owner accepts its tier once. Its routines arrive approved at that tier, and no higher.
-
v.
Removed
Uninstalling takes its tools, routines and tiles away together.
A skill holds up to twelve tools of up to eight steps each, guidance of up to 2,000 characters, six routines and six tiles.
III.
Your own tools
Your endpoint, called by your steward.
For the household that runs its own services. Name a tool, give it an address and the arguments it takes, and the steward can call it like any other, within its tier.
Address
HTTPS only
Each call
Signed with HMAC-SHA256 and a timestamp
Waits
Ten seconds for an answer
Reads
Up to 16 KB of the reply
Tier
T1 at the least, never a free read
Its secret
In the vault, used by that tool alone
IV.
Model Context Protocol
Other agents may ask. They may not act alone.
Castald speaks MCP in both directions: as a server that other agents can be granted, and as a client that mounts servers of your choosing.
Castald as a server
An agent asks for access; a member grants it on a consent page that names the agent, what it may do and the highest tier it may reach. Grants show under Connected apps and can be revoked at any time.
- Sign-in
- OAuth 2.1 with PKCE, dynamic registration and published metadata. A device code for agents without a browser.
- Tokens
- An hour for access; refresh tokens rotate and last thirty days.
- Tiers
- Capped by the member’s role. No agent ever holds T3, and T2 waits for a person.
- Spaces
- One grant may reach several homes and businesses, each under its own rules.
tools:read
tools:call
events:read
Servers mounted in Castald
Add the address of any MCP server and its tools become the steward’s, in that one home or business. Its token is kept in the vault.
- Reads
- Run freely, whether the server marks them or you do.
- Writes
- Ask first.
- Destructive
- Need the owner’s passkey.
- Described
- The steward is told each server’s name and what its tools do.
A steward for the whole estate.
Tell us a little about your house and your business. We will be in touch.