Pulso 🫀

An open-source, headless reliability backend built for agent-driven investigation.

Pulso's direction is to turn telemetry into actionable reliability signals with supporting evidence for external agent workflows, with raw queryable retention as an opt-in capability. The target combines bounded cross-signal service summaries with opt-in, budgeted model assessments; your workflow owns deeper investigation and policy-gated actions. Assessment, object-storage overhead and investigation have separate allowances. See the architecture for the target design and its boundary with current behavior.

Today, Pulso accepts logs, metrics, and traces over the protocols your collectors already speak (Loki push, Prometheus remote write, and OpenTelemetry), stores them as Parquet in an S3-compatible bucket, and lets agents investigate them through a native Model Context Protocol interface. Prometheus- and Loki-compatible query APIs keep existing dashboards working. Experimental periodic reliability assessments combine bounded cross-signal windows with opt-in OpenAI Decisions, durable MCP signals and capped external investigation handoffs. Processing-only ingestion is not implemented; raw records are retained indefinitely unless expiry is configured.

Experimental error tracking accepts Sentry SDK JSON errors into the same object storage and exposes error events and issue groups through MCP. An optional error monitoring index adds compact retained-history summaries and replayable changes without gating ingestion. Its first bounded detector emits measured error-rate increases with compact evidence for an external workflow to consume; it does not execute that workflow.

Why Pulso

• One system instead of three. Logs, metrics, and traces share one ingest path, one storage format, and one query surface.
• Object storage is the source of truth. No database, no local disks to back up, no consensus service. Nodes are disposable.
• Agent-native. Diagnosis tools are part of the server, not a wrapper bolted on afterwards, and they stay read-only by design.
• No UI. Bring your own dashboards, or your own agent.

Warning

Pulso is early software. Logs, metrics, and OTLP/HTTP traces work end to end. Native alerting and opt-in trace lifecycle policies are experimental; bounded Tempo lookup/search/discovery are experimental, while full TraceQL and deployed Tempo-client parity are not implemented yet. Bounded catalogs require explicit event-time retention; unlimited-retention catalogs remain unbounded. Expect breaking changes between releases.

Try it locally

You need mise, a stable Rust toolchain, and Docker.

mise install                  # Erlang and Elixir
docker compose up -d          # local S3-compatible storage
mix setup
mix phx.server

In another terminal, push a log line and read it back (local development accepts any tenant without a token):

curl -X POST localhost:4000/loki/api/v1/push \
  -H 'Content-Type: application/json' -H 'X-Scope-OrgID: demo' \
  -d "{\"streams\":[{\"stream\":{\"service_name\":\"api\"},\"values\":[[\"$(date +%s)000000000\",\"hello pulso\"]]}]}"

curl -G localhost:4000/loki/api/v1/query_range \
  -H 'X-Scope-OrgID: demo' --data-urlencode 'query={service_name="api"}'

The listener port can differ per checkout; mix phx.server prints the one it uses.

Documentation

The documentation covers deploying Pulso with the Helm chart or container image published with every release, configuration, sending telemetry, querying, limits, and monitoring Pulso itself. The architecture explains the design.

Contributing

Run mix precommit before opening a pull request; it runs the same compile, format, and test checks as CI. AGENTS.md describes the codebase conventions.

License

Pulso is released under the MIT License.