Changelog

What shipped, newest first: each node release (Android app and Linux node) by its tag, and the hub and website milestones in between. Only things that are released.

Atom feed for feed readers.

  1. Node v0.2.10

    A node in three commands

    Linux node, Android app

    • aigw-node pair, install and up take a machine from nothing to a serving node: pair with the hub, install the service (systemd on Linux, launchd on macOS), then add the chat models found on the machine and wait until the hub has the node online. Each command is safe to run again.
    • A "tools" setting on a chat model starts llama.cpp with its chat template, so tool calls work on the Anthropic, Responses and chat completions APIs.
    • The Android app starts serving when it is opened, also on TVs whose firmware blocks apps from starting themselves after an update.
  2. Milestone

    Anthropic Messages and OpenAI Responses APIs, tool calls

    Hub

    • Chat models on your nodes now also answer the Anthropic Messages API (/anthropic/v1/messages, streaming and tool use included), so clients built for it can point at the hub.
    • The OpenAI Responses API (/openai/v1/responses) works on the same models, stateless: every request carries its whole conversation.
    • Tool calls pass through to llama.cpp on every API, with limits on the number and size of tools, and streamed tool calls are indexed the way OpenAI SDKs expect.
    • The same keys, scopes, local-only rule, quotas and audit apply as on chat completions.
  3. Node v0.2.9

    Nodes stay on the LAN address that answers

    Hub, Linux node, Android app

    • The hub advertises only the LAN address of the interface that carries its default route (lan.allInterfaces restores all interfaces; an optional CIDR allow list narrows what is advertised), so nodes no longer spend a dial on an address they cannot reach.
    • When a node has fallen back to the public URL, its check for the LAN path backs off: the interval doubles after each round without an answer, from 2 up to 30 minutes, on the Android app and the Linux node.
  4. Milestone

    Free, with no paid tiers

    Website

    • The pricing page now says what is true today: the hub and the nodes are free to run, and there are no paid tiers.
  5. Milestone

    Website live, served by the hub

    Website, Hub

    • The website, documentation, compatibility registry and in-browser demo are public. The hub serves them itself for every path no API route claims.
    • Strict Content Security Policy with hashed inline scripts, precompressed files and atomic deploys with rollback.
    • Download links always point at the current stable release through /downloads/latest/.
  6. Node v0.2.8

    End-to-end sealing, chat models and signed checksums

    Linux node, Android app, Security

    • Nodes open end-to-end sealed requests (HPKE, per-node job keys); the hub forwards a sealed body without being able to read it.
    • The Linux node runs chat models through a managed llama-server and streams the answer back through the hub.
    • Release checksums are signed (SHA256SUMS.sig) and verified by the installer and the node before an update.
    • Automatic benchmarks wait until the device owner’s protections allow them (nothing loads while the node is paused).
  7. Milestone

    End-to-end sealing

    Hub, SDKs, Security

    • Clients can seal a request to the node that runs it. The hub routes, meters and audits the job but cannot read the prompt or the answer.
    • Client helpers for Go, Python and the Android library, with shared test vectors; they refuse to send without a pinned node key.
  8. Milestone

    OpenAI-compatible chat on your own devices

    Hub, Linux node, SDKs

    • POST /openai/v1/chat/completions runs on chat models served by your nodes, streamed or not, under the same keys, scopes and quotas as every other call.
    • Thin Python and JavaScript helpers on top of the official OpenAI clients.
  9. Milestone

    Trusted groups

    Hub

    • Accounts can share their devices with family and friends in a trusted group; your own devices are always tried first.
    • Invitations are capped per inviter, so one account cannot flood someone else’s invitations.
  10. Milestone

    Self-service accounts with passkeys

    Hub, Security

    • Sign in at /account with a passkey, with recovery codes for a lost device. No passwords.
    • Owner dashboard: create client keys, pair devices and see their status.
    • Operator controls for accounts and invitations at /admin/accounts.
  11. Milestone

    Signed releases and safer installs

    Hub, Linux node, Security

    • Every release ships a signed SHA256SUMS; the hub checks the signature when a release is uploaded, and install.sh verifies it again on the device.
    • The installer refuses downgrades and stops when a signature, checksum or version cannot be verified.
  12. Milestone

    Zero-config LAN and a trusted public registry

    Hub, Linux node

    • The hub listens on the LAN by default and announces itself over mDNS, so a Linux node can find it for pairing without typing an address. Discovered hubs are never trusted automatically: pairing checks the certificate fingerprint.
    • The public registry decides a device by its hardware probe fingerprint, and the operator marks which owners’ results count as trusted.
  13. Node v0.2.7

    Android start-up fix

    Android app

    • Release builds keep their local databases intact, which fixes a crash at start on phones.
  14. Node v0.2.6

    Check for updates on demand

    Android app

    • Settings > Version can check for an app update right away instead of waiting for the next scheduled check.
  15. Node v0.2.5

  16. Node v0.2.4

    Publisher keys survive restarts

    Android app

    • Publisher keys the hub offers to a device stay pending across restarts until the owner answers.
    • Offers lost on devices paired with 0.2.3 are recovered from the hub.
  17. Node v0.2.3

    TV home screen fix

    Android app

    • The TV home screen shows the recent jobs card again; the release build had crashed on the TV.
  18. Node v0.2.2

  19. Node v0.2.1

  20. Node v0.2.0

    Self-updating nodes

    Linux node, Android app

    • First release with self-update: nodes fetch new versions from their own hub, try them and roll back if the new version does not come up.
    • The hub serves installs and updates itself (one-line install for Linux nodes).
  21. Milestone

    Public registry, waitlist and packaging

    Hub

    • Public read-only registry (GET /public/v1/registry) and a waitlist endpoint, with a CORS allow-list and per-IP limits.
    • GET /v1/nodes is redacted for non-admin keys; /metrics on the public port is admin-only.
    • Distroless hub image, Docker Swarm stack and a deploy script that verifies the live commit.
  22. Milestone

    Server hub and outbound nodes

    Hub, Linux node

    • Go server edition of the hub: SQLite store, virtual keys (scopes, rate limits, daily quotas, budgets), audit and usage, admin dashboard.
    • OpenAI-compatible API: /openai/v1/models, /embeddings (image data URLs) and /chat/completions passthrough to a configured provider.
    • Nodes connect over an outbound WebSocket: no inbound port, works behind NAT and on mobile data.
    • Router: loaded models first, then the lowest measured latency, then the fewest in flight; failover on transport errors.
    • Compatibility registry: nodes file their benchmark results (works / works with tweaks / fails) automatically.
  23. Milestone

    Multi-model gateway, TV support and the client library

    Android app, SDKs

    • Model packages (manifest schema 1, SHA-256 verified), per-model scheduler with bounded queues, memory budget, lazy load and idle unload.
    • Two runtime builds: LiteRT 2 for phones (GPU / NPU / CPU) and LiteRT 1.4 + NNAPI for 32-bit TVs.
    • Fixed 9 s NNAPI runs on the TV; EfficientNet-Lite4 now runs in about 22 ms on its NPU. Per-request timing breakdown in every answer.
    • Same-device Binder IPC with a signature permission, and the Android client library (IPC, then hub, then app fallback).
    • mDNS discovery of other gateways, never trusted automatically.
  24. Milestone

    First prototype

    Android app

    • Single-model gateway on a Pixel 7 Pro: HTTPS with per-client tokens, LAN-only binding, rate limits.
    • CLIP ViT-L/14-336 (fp16 weights, 582 MB) on the GPU: about 1.5 s per photo, embeddings identical to the fp32 original (cosine 1.000).