Rapida Voice
A self-hostable orchestration platform for real-time conversational voice systems, integrations, and observability.
What does this agent do, and when should you use it?
Rapida is an open-source, end-to-end voice AI orchestration platform written in Go for teams that want control over deployment, credentials, and branding. It uses gRPC for bidirectional communication and organizes real-time audio, STT, TTS, VAD, telephony, and multi-channel integrations. Its Docker Compose deployment includes a UI, nginx API Gateway, Web API, Assistant API, Endpoint API, and Integration API, with optional OpenSearch and Document API knowledge services. The stated platform scope also includes agent state management plus call logs, streaming events, tool traces, latency breakdowns, metrics, and dashboards. It is best suited to agencies and enterprise teams willing to operate and configure a multi-service voice stack, rather than buyers seeking only a hosted voice endpoint.
With make up-all, Rapida launches its UI, nginx API Gateway, Web API, Assistant API, Endpoint API, and Integration API. The platform streams real-time audio over gRPC and, according to its documented features, orchestrates STT, TTS, VAD, models, prompts, tools, and telephony channels; credentials for providers such as OpenAI, Anthropic, Deepgram, and Twilio are configured in service YAML files. It produces call logs, streaming events, tool traces, and latency data, and make up-all-with-knowledge additionally starts OpenSearch and the Document API. For extension work, the repository directs new STT/TTS providers to api/assistant-api/internal/transformer/ and new telephony channels to api/assistant-api/internal/telephony/.
- An agency delivering white-label voice assistants that needs to retain customer credentials, branding, and infrastructure boundaries.
- An enterprise platform team deploying a real-time voice system with a UI, gateway, and several API services for internal operations.
- A backend team that needs to configure OpenAI, Anthropic, or custom inference alongside its own models, prompts, and tools.
- A telephony automation team adding a voice channel, starting from `api/assistant-api/internal/telephony/`.
- An operations team investigating call latency or tool execution through call logs, streaming events, and tool traces.
What are this agent's strengths and limitations?
- The documented Go and gRPC design targets low-latency, bidirectional real-time audio communication.
- One deployment includes a UI, gateway, and multiple specialized API services rather than only a single voice component.
- It explicitly supports self-hosted or managed operation and lets teams select OpenAI, Anthropic, open-source models, or custom inference.
- Its observability scope is concrete: call logs, streaming events, tool traces, latency breakdowns, metrics, and dashboards.
- The full Docker service set requires Docker, Docker Compose, and at least 16GB of RAM.
- Teams must configure API keys in YAML for whichever model, speech, or telephony providers they choose.
- Non-Docker local operation requires separately running PostgreSQL, Redis, and OpenSearch.
- Licensing needs verification: repository metadata says `NOASSERTION`, while the README claims GPL-2.0 with logo-retention conditions and describes a commercial license.
How do you install or deploy this agent?
Prerequisites: Docker, Docker Compose, and 16GB+ RAM for all services. Run:git clone https://github.com/rapidaai/voice-ai.git && cd voice-aimake setup-local && make build-allmake up-all
Then check the deployment with docker compose ps. The UI is available at http://localhost:3000; the nginx API Gateway is at http://localhost:8080. To include knowledge services, run make up-all-with-knowledge.
How do you use this agent?
After startup, edit the YAML configuration and add API keys for the chosen providers: docker/web-api/web.yml, docker/assistant-api/assistant.yml, docker/endpoint-api/endpoint.yml, docker/integration-api/integration.yml, and docker/document-api/config.yaml for knowledge services. The README explicitly gives OpenAI, Anthropic, Deepgram, and Twilio as credential examples. Use make logs-all for all logs, or make logs-web and make logs-assistant for individual services; rebuild changes with make rebuild-assistant or make rebuild-all. For a non-Docker run, use go mod download, go build -o bin/web ./cmd/web, and ./bin/web; PostgreSQL, Redis, and OpenSearch must be running separately.
How does this agent compare with similar options?
The README positions Rapida against vendor lock-in by allowing OpenAI, Anthropic, open-source models, or custom inference; it does not name specific competing products.