Integrations
Everything Marg plugs into is an environment line or a folder — no adapters to write. Models are on their own page: Models.
Observability
| OpenObserve (default) | The open-source default: traces, metrics and logs over OTLP/protobuf, with dashboards and alerts as code (npm run o2:bootstrap creates them idempotently). |
| Any OTLP backend | Marg emits OpenTelemetry in the GenAI semantic conventions (gen_ai.* on model and tool spans, agent_* metrics, logs that carry the trace id). |
| Langfuse | Langfuse ingests OTLP traces and maps the GenAI attributes to its generations view: model, tokens, cost per turn. |
| Datadog | Send OTLP to the Datadog Agent (OTLP ingest on 4318) or an OpenTelemetry Collector with the Datadog exporter; LLM Observability picks up the gen_ai.* spans. |
| Grafana Cloud (Tempo, Mimir, Loki) | Grafana Cloud's OTLP gateway takes traces, metrics and logs in one endpoint; the dashboards in observability/openobserve/ translate to Grafana panels query by query (npm run o2:dashboards prints them as JSON). |
| Honeycomb | Honeycomb takes OTLP/protobuf directly; one dataset per OTEL_SERVICE_NAME. |
| Arize Phoenix | Phoenix (open source) and Arize AX read OTLP traces with the GenAI attributes into their LLM trace and eval views. |
| SigNoz | Open-source APM on ClickHouse; traces, metrics and logs over OTLP. |
| New Relic | New Relic's OTLP endpoint (EU or US); AI Monitoring reads the gen_ai.* spans. |
| Jaeger (traces only) | Traces over OTLP; metrics and logs need another backend or a collector that fans out. |
Stores
| Redis (rate limits, caps) | Counters for rate limits and usage caps across gateway replicas. |
| MongoDB (sessions, ledgers) | Sessions, the usage ledger and events, durable across restarts and shared by replicas. |
Speech
| Deepgram (voice notes) | The widget's mic sends a voice note; the transcript lands in the composer for review before it is sent. |
| Whisper-compatible (OpenAI, Groq, self-hosted) | Any /audio/transcriptions endpoint: OpenAI, Groq, or a self-hosted faster-whisper / whisper.cpp server — the same provider branch. |
Decisions
| Laya and System One servers | The decision model in Verdict's cascade: routing, reading a confirmation, checking a condition, answered with calibrated probabilities in ~30 ms before an LLM is asked. |
Tools
| MCP tool servers (any language) | Your backend sits behind an MCP server the gateway spawns over stdio: TypeScript (the templates), Python (FastMCP), Go, Java, anything. |
| From another agent framework | Two seams. |
Hosting
| Render, Fly, Railway, Docker | One non-root image with a health check; Compose profiles for the stores, observability, decisions and the TLS edge; a Render Blueprint for a demo. |
Coding agents
| Claude Code, Codex and other coding agents | The method for building a domain ships as a skill a coding agent follows: the worksheet, the build order, the gate, the classification rule. |
Source: docs/integrations/README.md — edit it there; the site mirrors the repository.