Agents are agency, projected. We tailor the harness.

The replacement for general-purpose coding harnesses. Operator with socket and pipe I/O, fully customizable run-loop, tool surface, subagent taxonomy, and cognitive prompt (purpose / ego / fear). Systemd-supervised. Standard JSONL session format. Best-in-class steering and observability — debuggable, in-house, ours.

// maicraft.session
contract Session {
  in:   { harness: AST, tools: ToolDef[], prompt: Purpose }
  out:  { trace: JSONL, certs: Cert[] }
  pre:  identity.signed_by("maisec") && harness.admitted
  post: trace.hashed && certs.every(c => c.verifier == "lean4")
}

Monotonically narrowing token-mint engine over a meta-federation of major OAuth providers, unified into a single human-in-the-loop identity. Webauthn / passkey grant of standing agency to an instrumental agent. Vault-backed credentials. The fiduciary spine of the system.

// maisec.token.mint
contract Mint {
  in:   { subject: DID, scope: Scope, ttl: Seconds }
  out:  { token: SignedJWT, attestation: Cert }
  pre:  subject.webauthn_assented && scope <= subject.standing
  post: token.ttl <= scope.ceiling && revocable
}

Abstract Protocol Syntax Language — the contract dialect that lets agents write fiduciary-grade software. Every node declares (type_in, type_out, pre, post, complexity, parallelism). A graph-first compiler proves the application before any code is written. Agents fill typed holes, submit (impl, cert), the compiler admits or rejects. Admissible parts compose to an admissible whole by construction.

// maiapsl.node
node Settle {
  type_in:   { trade: Trade, ledger: Ledger }
  type_out:  { ledger: Ledger, receipt: Cert }
  pre:       trade.signed && ledger.balanced
  post:      ledger'.balanced && ledger'.delta == trade.amount
  complexity: O(log n)
}

A combined vector and graph database storing the full record of every agent operation — sessions, subdivisions, embeddings, scores, taxonomies. The substrate of iterative agent improvement. The benchmark surface. The audit trail.

A capability venn-diagram as dependency manager. Provisions OS, drivers, libraries, containers, and SIPs — agent self-improvement protocols covering memory, reasoning, browser emulation, credential management. Local package management becomes universal compute routing.

An LLM gateway sidecar that strips known prompt-injection payloads at the wire and escalates intent-injection attempts (volunteered credentials, role escalation, exfiltration requests) to the steering layer. Replaces the default gateway. Auditable. Configurable.

Refines any request into a falsifiable claim, decomposes it into a taxonomy of agents drawn from maivec, and executes that taxonomy with full auditing, observability, and benchmark. The "limbs" of an Operator agent — rapid, parallel, contained.

A federation of compute nodes registered against unspoofable hardware/network identity hashes. Every message between nodes is encrypted at send and decrypted at receive against a sender/receiver-derived key — messages outside the closed cycle are simply garbled and discarded. The final walled garden of TCP/IP security.

A git wrapper that enforces agent identity and binds a notes-database to every artifact. Inverts the broken claude-code execution model — observe, then act. Coordination for agent swarms in parallel without the trampling.

A React-Native client tunneled to your maigent over wireguard. Federates with telegram, groups, and topics — closes the desktop-client security hole and renders dynamic agent-built artifacts inline.

A signed registry of agent capabilities, version-controlled and trust-managed by role and group through maisec. Each cast carries a validation framework proving it delivers what it claims.

The instance itself: a secure, anonymous moltaicraft cloud installation with full moltai services integrated — telegram, sipcasts, dashboards, complete backups, and top-of-class security — ready to be granted standing agency by its principal.