Developer preview

Build agents around a decision contract, not a prompt.

T2K gives your application a persistent semantic layer for what exists, what is true now, what matters, what is allowed, what ran, and what happened next. You keep the model runtime and tools. T2K keeps the decision legible and governable.

Open the source Open the public schema

1 runtime dependency Node 20.10+ Apache-2.0 + DCO 24-episode golden demo verified

A real package boundary

Portable core. Operable control plane. Your product.

T2K does not require your agent, connector framework, or vertical workflow to move inside Studio. Each runtime remains independently testable, versioned, and replaceable.

01 / Core library

Meaning and contracts

Exact schema execution, typed contracts, pack compilation, reference replay, reward aggregation, semantic hashes, and the control-plane client.

@t2kai/core

02 / Studio

Governance and operations

Registry, graph ownership, evidence, Decision Contexts, authority, episodes, active policy delivery, held-out replay, human review, REST, MCP, and audit.

studio.t2k.ai

03 / Your system

Vertical behavior

Domain packs, business facts, model runtime, tools, connectors, execution, measurements, and the user experience.

your agent or product

Open developer preview

Start with one inspectable decision loop.

The standard, compiler, typed contracts, client, reference evaluator, conformance kit, synthetic Harborlight replay, and tested project scaffold are public under Apache-2.0. The provenance-backed npm packages and clean-install path are verified on Node 20 and 22.

01
Create locallynpx create-t2k@latest my-decision-loop
02
Run the computed replaycd my-decision-loop && npm start
03
Edit the contractontology + facts + policies + evidence
04
Keep authority explicitrecommendation != authorization

my-decision-loop/generated project

ontology-pack.json
decision-context.json
policies/
  baseline.json
  candidate.json
episodes/
  holdout.json
src/
  run.mjs

$ npm start
replay.status              passed
candidate.action           rebalance_route
recommendation.status      eligible_for_human_review
authorization.status       not_authorized

Canonical agent loop

Separate understanding, deciding, acting, and learning.

Each boundary creates an object you can inspect later. No outcome silently becomes truth, reward, or a production policy update.

01Compile stateOntology + accepted claims + provenance + time
02Run policyVersioned model, prompt, rules, tools, and hashes
03AuthorizeNamed authority, risk, limits, and conditions
04ExecuteIdempotent attempt, external identity, and receipt
05ObserveSourced measurements, baselines, and windows
06EvaluateReward vector, guardrails, and attribution
07PromoteHeld-out replay, independent review, one active version, rollback

Three compatible interfaces

Use the transport your runtime already speaks.

TypeScript

Typed orchestration

T2kClient covers registry, graph, decision, execution, learning, publication, and rollback workflows. @t2kai/core@0.1.0 and create-t2k@0.1.0 are published with GitHub Actions provenance and verified registry attestations.

Inspect the repository

REST

Server-to-server control

Approved Studio integrations call versioned /api/v1 routes with scoped API keys. Never ship keys in browser code.

Review the integration path

MCP

Agent tool discovery

Connect at POST /api/mcp. MCP obeys the same scope, graph, mount, authority, and lifecycle controls.

See the closed loop

livenessGET /api/health

readinessGET /api/ready

schema/schemas/t2k-ontology-pack.v1.schema.json

releaseatomic activation + rollback

Start with one decision

Give your agent a world model and a receipt.

Compile a small pack, bind a private graph, and close one measurable episode end to end.

Clone the core Run the tutorial