Colloq v0.1Evidence v0.2 · two-process Iroh · open source →

One typed conversation for distributed AI.

Vantar builds Colloq, an open-source language for AI systems, services, and accelerators that coordinate through one typed, verifiable conversation. We also build Coldstack, the retrieval layer for robot data.

Built on the stack teams already run

RustQUICIrohAutomergeMirenMCAPROS 2PyTorch

Featured · Colloq

Open source

A programming language for machines that coordinate machines.

Colloq replaces separate client programs, server programs, and disconnected protocol definitions with one typed conversation graph. The compiler projects that agreement into compatible endpoint plans while preserving one inspectable identity across the data center.

One global program

The conversation is the computation.

Typed choices, streams, cancellation, and failures project into one checked local machine per role.

Authenticated data plane

Meaning survives the network.

Persistent Iroh identities and exact role/plan policy let independent servers execute the same identified Colloq Wire.

Machine-authored change

Evolution stays governed.

Automerge sync runs as an authenticated Colloq conversation, then passes through conflict, validation, and promotion gates.

58

Tests passing

4

Transports

2

Wire encodings

v0.1

Research prototype

Colloq

Research · v0.1

The language servers speak to think together.

Colloq begins with one typed conversation instead of separate client and server programs. Its compiler projects that graph into compatible endpoint plans, then preserves the same semantics across memory, TCP, authenticated QUIC, or Iroh. Automerge protects collaborative draft promotion, while Miren packages the current multi-node testbed.

Illustrative Colloq
conversation Generate(prompt: Prompt) -> stream<Token> {
    roles gateway, router, expert[*]

    gateway -> router: prompt within 2ms

    choice router {
        cached { router -> gateway: CachedResult; end }
        infer(expert) {
            router -> expert: prompt
            expert -> gateway: stream<Token>
        }
    }
}

Coldstack

Design partners

Robot data belongs on object storage.

Fleets generate 0.5–2 TB per robot per day. Coldstack keeps your raw MCAP in your own S3 bucket, builds a compact index, and answers one composable query — semantic, time-series, and metadata together. The retrieval layer for robot data.

Python
import coldstack

ns = coldstack.Client(api_key=...).namespace("fleet-a")

results = ns.query(
    text="gripper slipping on transparent object",   # visual semantic
    signal="torque_z > 5 for 2s and velocity < 0.1",  # time-series pattern
    filter={"robot_id": ["r-204", "r-207"]},          # metadata
    limit=50,
)

≥5×

Cheaper (target)

vs hot-storage stacks

1–3%

Index footprint

of raw log size

<100ms

Warm query

hot-namespace target

0 bytes

Raw data moved

stays in your bucket

Working engine, validated on real public robot logs — index footprint measured at 0.02% on a camera/lidar recording. In design-partner recruitment; not yet benchmarked at fleet scale.

Why Vantar

The tooling hasn't kept up. We build the missing pieces.

Robot data has no retrieval layer

Fleets generate 0.5–2 TB per robot per day. ~99.9% is never read again, but the failures and edge cases that matter need fast search across the entire corpus. Hot storage is ruinous, Glacier is unsearchable. Coldstack makes petabyte-scale robot logs searchable on object storage.

Distributed AI has no shared language

Models, tools, memory, accelerators, and services are still stitched together with separate APIs and deployment files. Colloq starts from one typed conversation and compiles compatible endpoint plans for every participant.

Distributed AI needs
one shared language.

Colloq is open source and early. Tell us about the workload you want to coordinate, or the robot logs you need to search.

Early access · Colloq design partners & Coldstack

We store your address to contact you about early access. Nothing else.