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

jellyfin/jellyfin

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

371 open issues

#17383 Which algorithm is used to create a collection poster?
#17380 Decouple `StartupWizard` from Initial Server Configuration
#17379 Missing-episode reconciliation creates duplicate phantom episodes when a series' aired and DVD/TMDB episode counts differ
#17372 Allow plugin tuner and guide provider types to supply their own configuration UI
feature
#17366 UpdateItem API endpoint throws ArgumentNullException on 'source' parameter when updating only 'tags' field
bug
#17364 automount hostname lookup .ignore
bug
#17363 Ubuntu 26.04 + Docker + NVIDIA GTX 980: Jellyfin playback always fails (0:00 / Fatal player error)
bug
#17357 Performance degradation for `/UserItems/Resume?userId=UUID...` API endpoint
bug
#17352 SyncPlay sessions persisting even after everyone closes their session
bug
#17351 SyncPlay not working in 10.11.11
bug
#17346 Trickplay tiling fails on specific aspect ratio content on AMD GPUs
bug
#17345 ScheduleDirect Image Limit Handling
bug
live-tv
#17341 Concurrent subtitle requests return corrupted subtitle files or crash with NullReferenceException
bug
#17332 strm-backed (remote HTTP) VOD library items get no reconnect-on-drop — reconnect flags only apply to Live TV (RequiresLooping)
#17319 Live TV: one slow/stalled tuner `Open()` freezes ALL Live TV for 100s (global lock held across the network connect); the shared-stream path has no connect timeout and leaks on failure
#17318 Jellyfin's handling of terminating TCP connections causing issues for STRM playback proxied from Dispatcharr
bug
#17316 Jellyfin not selecting the forced subtitle when the subtitle mode is "Always Play" and the audio and preferred subtitle languages do not match
bug
#17312 HEIC images are not indexed in "Home Videos and Photos" library although bundled FFmpeg decodes them
bug
#17301 Successful login immediately reverts the user's LastLoginDate (stale entity overwrite in LogSessionActivity)
#17296 Runaway Trickplay, 500+GBs of black jpg files in several random folders in appdata\local\temp\jellyfin
bug
#17295 Playback / No sound when specific audio tracks are selected, under a specific quality
bug
#17294 Trickplay generation runs away indefinitely on files with corrupted Duration metadata
bug
#17284 [12.0.0-RC2] TV library played/unplayed filter not working (movies filter correctly)
bug
#17279 Remuxing causing skipping within the file, but not affecting the duration within the file
bug
#17277 BDMV + TrueHD (AC3 core): PGS subtitle burn-in fails — stream index offset causes wrong track or "matches no streams"
#17270 [12-rc2] Series images under Recently Added on clients are incorrect
bug
#17267 [v12.0RC2] Subtitles do not appear properly if they're not UTF-8
bug
#17264 Subtitle search (RemoteSearch/Subtitles) returns empty for BluRay/DVD/ISO format videos
#17259 [12-RC2] Refresh guide takes very long and locks up Jellyfin
bug
#17256 SQLite foreign key constraint failures during large library scan with concurrent metadata and segment tasks
bug