Operate is a self-hosted AI SRE: an AI agent platform that detects, investigates, and resolves production issues. It runs a four-agent pipeline (Context, Root Cause, Verification, and Resolution) that connects to your logs, database, and codebase, then produces an evidence-backed root cause and a draft pull request for human review.

An AI SRE is an AI system that does the work of a site reliability engineer: watching production for errors and regressions, investigating alerts, finding root causes, and proposing fixes. Operate is a self-hosted AI SRE, so it runs inside your own infrastructure with read-only access and a mandatory human approval gate.

How does Operate find the root cause of a production issue?

When triggered from chat (Slack, MS Teams, and other tools) or by an automated alert, Operate pulls traces from log providers like Datadog, CloudWatch, GCP Cloud Logging, and SolarWinds, queries your MongoDB, PostgreSQL, or MySQL database in read-only mode, and reads your codebase via git including blame and diff. A Context Agent gathers signals, a Root Cause Agent proposes a cause with a confidence score, a Verification Agent independently checks the reasoning with a different model, and a Resolution Agent drafts a fix as a pull request.

Does Operate ship code automatically without a human?

No. Operate stops at a draft pull request. A human engineer must review and merge; nothing ships without approval. Every agent step is auditable so reviewers can see the evidence and reasoning behind each finding.

How is Operate deployed?

Operate is deployed as a Docker image into your existing infrastructure with no new cloud accounts. It runs on port 8080 under the /operate path, requires MongoDB and Redis, and uses an App ID from operate.bettrsw.com. It connects via read-only adapters to logs, databases, and observability stacks, and clones your repo at startup via an init container so agents have full git history.

Which integrations does Operate support?

Operate's context layer spans 60+ integrations across 14 categories: source control (GitHub, GitLab), logs (Datadog, CloudWatch, GCP Cloud Logging, SolarWinds), databases (MongoDB, PostgreSQL, MySQL, read-only), error tracking (Sentry, Rollbar), APM (New Relic, Grafana, OpenTelemetry), CI/CD (GitHub Actions, Jenkins, Vercel), feature flags (LaunchDarkly, Statsig), cloud infrastructure (AWS, GCP, Azure, Kubernetes), queues (Kafka, SQS, RabbitMQ), incident management (PagerDuty, incident.io), issue tracking (Jira, Linear, Asana), support (Zendesk, Intercom), status pages, and docs (Confluence, Notion). Triggers: Slack, MS Teams, and other chat tools for on-demand RCA, plus the observability stack for proactive monitoring. AI providers: Anthropic Claude, OpenAI/Azure Codex, enterprise models, or self-hosted open-source LLMs.

How does Operate avoid hallucinated root causes?

Operate uses a two-model verification step. The Root Cause Agent proposes a cause with specific evidence; the Verification Agent, running on a different model to avoid correlated errors, independently evaluates the reasoning. Speculative or unsupported conclusions are rejected before reaching the team.

Why not just point an LLM at our logs ourselves?

An LLM has no inherent knowledge of your logs, schemas, services, or known-bad patterns. Operate provides the connectors, permissions, query layer, repo context, multi-agent verification, and operational runtime that turn a generic model into a reliable production-investigation system.

Engineering teams that lose 30–40% of capacity to unplanned investigation: support escalations, error alerts, performance regressions. With Operate, support and product teams can trigger an RCA from Slack, MS Teams, or whatever chat tool they use, and only loop in engineers when a decision needs to be made.

← All repos

github

nats-io/nats-server

Pick an open issue and Operate will investigate its root cause.

445 open issues

#8396 Export/import - Reply to $JS.API.CONSUMER.RESET..... NOT crossing account boundary when sent by consumer
defect
#8393 Allow verify_and_map to derive identity from any certificate attribute, not just email/SAN/DN
proposal
#8388 /varz reports stale JetStream config.max_memory/max_storage after max_mem_store/max_file_store reload
defect
#8387 Leafnode sourcing with OptStartTime does not source messages in the past
defect
#8368 jetstream sequence inconsistent with actual message order
defect
#8367 Hub IPs are leaking to leafnodes even if no_advertise is set to true
defect
#8363 Flaky TestRegisterAndUnregister due to asynchronous ATS goroutine shutdown
defect
#8352 Allow stream, consumer & account names to contain arbitrary characters by encoding pathnames
proposal
#8350 Jetstream consumer createOrUpdate context deadline exceeded
defect
#8348 `JetStream meta: orphaned R1 stream assignments for an account with JetStream disabled ($G) cause a permanent /healthz 503 "not current: stream not found"`
defect
#8346 JetStream sourced stream reports lag=0 after source stream recreation but no longer ingests new messages
defect
#8339 Duplicate delivery to non-queue subscribers when a leaf-node cluster is splayed across a supercluster
defect
#8338 Service-import → stream-export chain drops messages when publisher and subscriber are on the same server
defect
#8325 Batch fast-batch JetStream proposals at the stream-to-RAFT boundary
proposal
#8322 JetStream dynamic MaxStore shrinks after restart because it is recomputed from current free disk (Bavail), causing previously valid stream limits to fail
good first issue
defect
#8320 JetStream follower quorumPaused can stall catchup and prolong recovery under load
defect
#8319 Server can fail to ever sync up after falling behind in clustered jetstream
defect
#8309 Cluster route TLS: FQDN from route URL is resolved to IP before TLS dial, causing x509 IP SAN verification failure
defect
#8299 Support using multiple disks
proposal
#8274 JetStream push consumers can deliver to subjects creator cannot publish
defect
#8271 Service-import request metadata can exceed max_payload
defect
#8270 Client TLS pinned-cert reload validates existing clients against old pins
defect
#8249 Jetstream: msgs & __msgs__
defect
#8171 Data race detected when using `tieredReservation` (CreateOrUpdate api functions)
defect
#8169 JetStream: periodic ~15s delivery gaps on sparse-filter pull consumer with max_ack_pending=1 (multi-consumer cluster, light load)
#8150 Jetstream does not recover after enabling TLS on cluster
defect
#8137 Small addition to the ClientAuthentication Interface
proposal
#8095 SOCKS5 proxy support
proposal
#8090 Allow to use upgradeWebsocket hander in an golang app embedding NATS - export handle function of server/websocket.go#1327
proposal
#8084 healthz should not report jetstream account errors for expired accounts
proposal