Realtime inference mesh

Stream every token.
Route every model.

Naga Bonar is an edge-native control plane that streams tokens, routes across model providers, and observes every request from a single low-latency mesh.

38msp50 first token
99.98%mesh uptime
40+models routed
mesh://stream · live
$ curl -N nagabonar.dev/v1/stream \
  -H "Authorization: Bearer ***" \
  -d '{ "route":"auto", "prompt":"..." }'
← streaming
data: {"token":"Naga"}
data: {"token":" Bonar"}
data: {"token":" mesh"}
routed → provider-3 · 38ms
Routes acrossOpenAIAnthropicMistralLlamaGeminiGroqlocal
Why a mesh

One control plane. Every model, every region.

01 · ROUTING

Auto model routing

Score latency, cost and quality per request. Fail over across providers with zero downtime.

prompt → [ mesh ] → best-of-N
02 · STREAM

Token streaming at the edge

SSE and WebSocket streaming from the nearest node — first token in tens of milliseconds, anywhere.

03 · CACHE
6.4×

Semantic cache hit-rate on repeat prompts.

04 · GUARD

Guardrails

PII redaction, rate limits and per-key budgets built in.

05 · OBSERVE

Trace every request

Tokens, cost, route and latency logged and replayable — no extra SDK.

06 · EDGE

Edge-native

Runs in 20+ regions. Deploy a route, it is global in seconds.

From key to stream

Three lines to a live token stream.

get a key

Join the beta, mint a scoped key with its own budget.

point at the mesh

Swap your base URL. Keep your existing SDK.

stream

Set route:"auto" and the mesh picks the fastest healthy provider.

# python
from nagabonar import Mesh
mesh = Mesh(key="nb-***")

for tok in mesh.stream(
    route="auto",
    prompt="explain the mesh",
    budget=0.02,
):
    print(tok, end="")
# first token in ~38ms, auto-routed
Private beta open

Put a mesh between you and every model.

Get a key, point your base URL at Naga Bonar, and stream from the fastest provider in every region.