Four layers. Fifty primitives.
Each repo lands on one layer. Ember = flagship.
Three reasons SaaS loses for AI infrastructure.
Audit, switching cost, pace of change.
SaaS works for stable problems; the 2026 agent stack isn't one. Five infra acquisitions in six months — Helicone, Langfuse, Promptfoo, Traceloop, Permify — say the layer is still settling. Three tests decide it:
The bricks are shipped. Here's the cube they build.
The primitives below are the parts. These 15 repositories wire them into a governed, eval-gated cube that runs end‑to‑end — the exact architecture the whitepaper describes, in TypeScript you can read. Published 2026‑05‑28, Apache‑2.0.
create-cube generator + cube.config manifest + CLI.bun start runs the full stack live. Public hub linking all 15.Read the world cleanly.
Stream parsers, document ingestion, pre-flight filters.
<think>...</think>). Three-language port.Every agent has a name & a signature.
DIDs, capability VCs, signed actions, attestations.
NIST and OWASP both published in 2026 that traditional IAM is inadequate for agent identity. These primitives implement the new model: per-agent DIDs, capability verifiable credentials, signed action receipts, and provenance-aware MCP servers.
The gateway, the sandbox, the throttles.
Substrate-level routing, rate limits, prompt injection, sandboxed execution.
shack_* meta-tools surface instead of dumping every downstream schema (~5 k tok each, breaks at ~50 servers). Security pipeline on every call: sandbox, allow/deny, hooks, redaction.The eval harness - the flagship.
Quality-gated SOP execution with independent LLM evaluators.
An eval-first approach: every SOP execution gets independent multi-dimension scoring before it ships. Compare to Inspect (UK AISI), Promptfoo (now under OpenAI), and DeepEval - all OSS. Skip Braintrust ($80M-funded SaaS) unless you specifically need their UX.
Every action replayable.
Byte-deterministic transcripts, reproducibility seeds, MCP-backed session capture.
OSS-first observability stack pairs well with Langfuse, Arize Phoenix, OpenLLMetry, Logfire free tier - all open or OTel-native. Skip Datadog LLM Obs unless your firm already pays Datadog for everything.
Content-addressed everything.
CIDv1 + Ed25519 + DID manifests for every artifact.
For the rest of the data fabric, the OSS-first stack is pgvector + DuckDB + Qdrant + Apache Iceberg. Add Memgraph or Neo4j Community for graph. No paid vector DB needed under ~50M vectors.
The signed knowledge network.
Trusted teams publish playbooks; agents retrieve with cryptographic verification.
./LEXICON.md + global ~/.claude/LEXICON.md with Applies when: matching. Three modes, never derails.Multi-agent coordination, working memory, tool calling.
The cognition substrate without the framework lock-in.
For the agent framework itself, the OSS picks are LangGraph, Mastra, Letta, Pydantic AI, smolagents. These below are the orthogonal primitives that any framework benefits from.
The write side - payments, publishes, fleets.
Durable agent tasks, agent-to-agent payments, autonomous release pipelines.
For durable workflow execution, pair with OSS engines: Temporal, Restate, DBOS, Inngest OSS, Hatchet, Trigger.dev OSS.
The bus, the transport, the schemas.
Type-safe event bus, encrypted agent RPC, federated Q&A.
Pair with OSS protocols: MCP (Apache 2.0, Linux Foundation), A2A (Apache 2.0, Linux Foundation), CloudEvents (CNCF), AsyncAPI, Apicurio Registry (Apache 2.0, the only fully-OSS multi-format registry).
on, once, off, emit, onAny, waitFor). Node 20+, browsers, Bun, Deno, edge.Memory that survives agent rotation.
Portable schema. Episodic, semantic, procedural - one taxonomy.
For the memory engines themselves: Letta (Apache 2.0), Cognee (MIT), Graphiti (Apache 2.0) - all OSS. Skip Zep Cloud if Graphiti standalone covers you.
Same primitive, three languages.
TypeScript for the team that ships fast. Python for the team that ships science. Rust for the team that ships forever.
Most of these primitives ship as a coordinated trio - the TypeScript version, the Python version, and the Rust version - published from a single agent-ports meta-repo. Agents and humans pick whichever language fits the surface they live on. The wire formats stay identical across all three so cross-language deployments work without translation.
TypeScript
For agent UIs, web-facing tools, MCP servers run inside Node / Bun / Deno, and integration glue. The first port for each primitive.
Python
For ML pipelines, data work, anywhere the science already lives. Same wire format, same semantics.
Rust
For the substrate edge - gateways, sandboxes, validators. Where performance and safety both matter.
Vollko is a small senior team. We ship the OSS primitives above because the AI-native organization we want to help firms build needs them - and a primitive that is not open cannot be a foundation.
Build the AI-native firm