Tracer-Cloud/opensre

What it claims

OpenSRE is in public alpha. The page says core workflows work for early exploration but are not yet stable, and APIs and integrations may change. Today the shipped product is the agent runtime. The RL training and evaluation environment is stated as the mission, and the concrete piece that exists so far is a suite of cloud-backed end-to-end failure tests in tests/e2e. The page gives no scenario count and no benchmark results. The licence is Apache 2.0.

How it works

When a question or alert arrives, the agent:

1. Gathers correlated logs, metrics, traces and recent deploys.
2. Optionally masks sensitive identifiers before calling an external LLM.
3. Tests hypotheses in a tool-calling loop across the connected systems.
4. Returns an evidence-linked answer with suggested next steps.
5. Optionally executes remediation.
6. Posts a summary to Slack, PagerDuty or Telegram.

flowchart LR
  A[Alert / question] --> B[Fetch context]
  B --> M{Mask IDs?}
  M --> L[External LLM]
  L --> T[Tool-calling loop]
  T -->|next hypothesis| L
  T --> R[Evidence-linked answer]
  R --> X{Remediate?}
  X -->|optional| E[Execute action]
  R --> P[Slack / PagerDuty / Telegram]

The page says the old graph and chain framework layers were removed. The current architecture is described in AGENTS.md, and contributors should start from main.py.

How to use it

Install. The installer pulls the latest build from main, not from a tagged release, and does not need sudo:

curl -fsSL https://install.opensre.com | bash -s -- -gh

& ([scriptblock]::Create((irm https://install.opensre.com/install.ps1))) -gh

Running opensre with no arguments validates your account and opens a REPL, which requires a TTY. The shell only opens for an active account, and the first launch activates the hosted model. The REPL has slash commands such as /cost, /resume, /compact, /integrations verify and /agents.

Headless mode, for scripts and CI. It supports stdin prompts, JSON output and tool approvals:

opensre ask "why is checkout-api slow?"

Python API. This requires a source checkout:

from bootstrap.embedded import start_embedded_session
session = start_embedded_session()
result = session.chat("why is checkout-api slow?")
if result.answered:
    print(result.primary_response_text)

Other commands: opensre integrations setup, opensre fleet scan, opensre update, and opensre uninstall, which removes all local data.

Deployment options:

• OpenSRE Cloud managed gateway: an organization admin provisions it in the web app.
• AMI + systemd: make build-gateway-image then make deploy-gateway. This path supports the Telegram gateway only and uses no Docker.
• Self-hosted container (Railway, ECS or Vercel): deploy the repo Dockerfile, set MODE and LLM_PROVIDER, and supply that provider's API key. Add DATABASE_URL (Postgres) only for shared state, for example Slack Events API replicas.

Limits

• Alpha status: APIs and integrations may evolve.
• The default CLI path is tied to a hosted account and model. The page does not say whether the local CLI can run fully offline or bring-your-own-key without signing in. Only the self-hosted container path documents LLM_PROVIDER plus a key.
• Telemetry is on by default. First-party product analytics and Sentry error reporting are opt-out. To disable both, set export OPENSRE_NO_TELEMETRY=1.
• Identifier masking before external LLM calls is optional, not the default. Check this before connecting it to production data.
• Remediation can execute actions. The page names tool approvals for headless mode but does not describe the guardrails in detail.
• The SWE-bench-style benchmark has no published scenarios, metrics or results yet. Treat it as roadmap, not something to evaluate against.