Speko Gateway

Speko Gateway is the open customer-side runtime for real-time voice AI. It gives agents one local streaming protocol across voice providers, keeps BYOK credentials inside your process, and can optionally use Speko for managed routing, observability, and consolidated billing.

Early preview: the protocol is versioned, but breaking changes may occur before the first stable release.

Versions and changelog

Pin a released version in production. Each release has notes on the releases page. Watch the repo with Custom → Releases to get notified.

Tag	What it points to	Use it for
0.1.0	One exact release. The workflow refuses to rebuild a released version.	Production, when you want to choose every upgrade.
0.1	The newest 0.1.x release. Gets fixes only.	Production, when you want fixes automatically.
latest, edge, sha-<commit>	Every merge to main. Moves several times a day.	Testing unreleased changes. Not for production.

Add Gateway to a Pipecat agent

Gateway includes native Pipecat STTService and TTSService processors with utterance commits, streaming sentence synthesis, barge-in cancellation, and a Pipecat Cloud-compatible sidecar startup hook. See the complete Pipecat integration guide.

Add Gateway to a LiveKit agent

The public image contains both the Gateway binary and its Python integration. Apply this diff to the official LiveKit Python agent starter Dockerfile:

 ARG PYTHON_VERSION=3.14
+FROM spekoai/gateway:0.1 AS speko-gateway
 FROM ghcr.io/astral-sh/uv:python${PYTHON_VERSION}-bookworm-slim AS base

 FROM base AS build
 WORKDIR /app
 COPY pyproject.toml uv.lock ./
 RUN mkdir -p src
 RUN uv sync --locked
+COPY --from=speko-gateway /opt/speko/python /opt/speko/python
+RUN uv pip install --python /app/.venv/bin/python /opt/speko/python
 RUN uv run --module livekit.agents download-files
 COPY . .

 FROM base
 ARG UID=10001
 RUN adduser \
     --disabled-password \
     --gecos "" \
     --home "/app" \
     --shell "/sbin/nologin" \
     --uid "${UID}" \
     appuser
 COPY --from=build --chown=appuser:appuser /app /app
+COPY --from=speko-gateway /usr/local/bin/speko-gateway /usr/local/bin/speko-gateway
+RUN install -d -o appuser -g appuser -m 0700 /run/speko
 WORKDIR /app
 USER appuser
-CMD ["uv", "run", "src/agent.py", "start"]
+CMD ["sh", "-c", "/usr/local/bin/speko-gateway & exec uv run src/agent.py start"]

Then select Speko for the voice legs where the agent creates its AgentSession:

from livekit.plugins import openai

from speko_gateway.livekit import STT, TTS

session = AgentSession(
    stt=STT(
        language="en",         # default
        provider="auto",       # default; set when several BYOK STT keys exist
        model="auto",          # default: the provider's catalog default
        credential_source="auto", # or "byok" / "managed"
        sample_rate=16_000,    # default
    ),
    llm=openai.LLM(model="gpt-4.1-mini"),  # any LiveKit LLM plugin
    tts=TTS(
        provider="auto",       # default: the configured BYOK vendor, or the managed plan's pick
        model="auto",          # default: the provider's catalog default
        voice="",              # default: request, configured fallback, or catalog default
        language="en",         # default
        sample_rate=24_000,    # default
        credential_source="auto", # or "byok" / "managed"
    ),
)

With a Speko API key the LLM can come from Speko too, served by the hosted Router (router.speko.dev):

from speko_gateway.livekit import LLM, STT, TTS

session = AgentSession(
    stt=STT(),
    llm=LLM(
        provider="auto",          # default: relay picks; set with model= to pin a route
        model="auto",             # default: GET router.speko.dev/v1/models lists the options
        objective="balanced",     # default: or "quality", "latency", "cost"
        max_output_tokens=8_192,  # default
        reasoning_effort=None,    # default: provider default; or "none" ... "xhigh"
    ),
    tts=TTS(),
)

reasoning_effort is the Router's reasoning control. Use "none" to turn reasoning off. "minimal" asks for the least reasoning the routed model allows (the Router maps it onto each provider's own scale), and "low", "medium", "high" and "xhigh" ask for more. Unset, no field is sent and the provider default applies. Any other value raises ValueError when the LLM is constructed. While reasoning is enabled (any value but "none") the Router refuses temperature and top_p.

"none" is accepted on models that cannot reason, where it changes nothing, except on older Routers, which refuse it there (e.g. on gpt-4.1). Four models always reason and refuse "none"; "minimal" is their floor: Inception mercury-voice, gpt-oss-120b on Cerebras and Together, and Together zai-org/GLM-5.3-Flash. Every other value, "minimal" included, needs a model that can reason: a pinned model that cannot is refused (capability_unsupported), and auto routing considers only models that can.

STT accepts transcription options in the same canonical vocabulary as the hosted Speko API, plus a per-vendor passthrough for each provider's own settings:

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗