OpsClaw is a self-hosted SRE agent for one operations team. It continuously checks infrastructure, investigates alerts, proposes or performs remediation within explicit policy, and keeps the evidence needed to understand what it did.

It runs as a daemon beside the systems it watches, not as a SaaS service. The daemon provides an installable web UI, an operator CLI, optional Telegram and webhook channels, durable SQLite state, and team memory through Hindsight.

What it does

• Schedules proactive scans and correlates findings into incidents.
• Operates across SSH, local, and Kubernetes targets, with typed access to service resources.
• Applies autonomy per project, environment, or target: observe, suggest, act_on_known, or auto.
• Queues risky changes with the exact command and context needed for approval.
• Routes every target command through one policy choke point with e-stop, cost, rate, scope, and audit enforcement.
• Records a tamper-evident audit chain and produces incident postmortems.
• Gives the team a shared Operations chat and a durable conversation per project environment. Environment chat is also a tool-authorization boundary.
• Tracks combined OpsClaw and Hindsight LLM spend against a daily cap.

Standard installation

The supported standard install is Docker Compose. It runs OpsClaw with the pinned Hindsight memory service and stores config, SQLite data, memory, and prepared target credentials in named volumes.

Requirements: Docker with Compose, an OpenRouter API key, and a host from which the intended targets are reachable.

git clone https://github.com/DCPRevere/opsclaw.git
cd opsclaw
cp deploy/.env.example deploy/.env
mkdir -p deploy/creds
$EDITOR deploy/.env

Set HINDSIGHT_LLM_API_KEY, OPENROUTER_API_KEY, and OPSCLAW_MODELS in deploy/.env. Pin OPSCLAW_VERSION to a release tag for production.

Onboard before starting the daemon. A fresh daemon cannot boot in live mode until onboarding has generated its gateway secret.

docker compose -f deploy/docker-compose.yml --env-file deploy/.env \
  run --rm --no-deps opsclaw-daemon \
  opsclaw-cli onboard --hindsight-url http://opsclaw-hindsight:8888

docker compose -f deploy/docker-compose.yml --env-file deploy/.env up -d

docker compose -f deploy/docker-compose.yml --env-file deploy/.env \
  exec opsclaw-daemon opsclaw-cli doctor

Hindsight can take roughly 40 seconds to become ready on its first start. If doctor reports only Hindsight as unavailable, wait and run it again.

The web UI listens on http://127.0.0.1:8484. Keep it private: use localhost, a VPN/tailnet, an SSH tunnel, or a TLS reverse proxy. Do not expose the daemon directly to the public internet.

Target credentials

Place host-provided credentials in deploy/creds/. The Compose credential initializer copies them into a private volume, assigns the daemon UID, and applies restrictive modes. Reference the resulting path as /home/opsclaw/creds/<name> while adding a target.

docker compose -f deploy/docker-compose.yml --env-file deploy/.env \
  exec opsclaw-daemon opsclaw-cli config projects add

A local target in Docker means the OpsClaw container, not the Docker host. Use SSH when the host itself is the intended target.

Operating model

Configuration follows this hierarchy:

Project
└── Environment
    └── Target
        └── Resource

Policy attributes inherit down the hierarchy; credentials never do. A target is one connection and audit identity such as shop::prod::api. Resources add service-specific access such as PostgreSQL, Prometheus, Argo CD, RabbitMQ, or Elasticsearch without weakening the owning target boundary.

Team chat follows the same model:

• Operations is the global coordination conversation and cannot invoke target tools.
• Each project environment has a shared conversation whose transcript and tools are restricted to that environment.
• Every operator sees the same history and can continue another operator's work. There are no private operator transcripts.

The CLI selects an environment conversation explicitly:

opsclaw-cli chat --remote --conversation shop::prod
opsclaw-cli chat --remote --conversation shop::prod --history

Safety

OpsClaw treats model output as a proposal, never as authority. Tool arguments are resolved against configured targets, command risk is classified in code, and target execution is reachable only through policy enforcement.

Useful controls:

opsclaw-cli estop                         # local, works if the daemon is wedged
opsclaw-cli approvals list               # inspect exact queued commands
opsclaw-cli blocked-actions              # mutations refused in observe mode
opsclaw-cli audit verify                 # verify the audit chain
opsclaw-cli usage                        # spend, cap, and remaining headroom
opsclaw-cli status                       # remote daemon and active-turn status

Start new targets at observe, inspect blocked-actions, then promote their autonomy deliberately. Use dedicated, least-privilege credentials that can be revoked without affecting a human operator.

Common commands

opsclaw-cli doctor
opsclaw-cli config list
opsclaw-cli incidents list
opsclaw-cli incidents ack <id>
opsclaw-cli incidents resolve <id>
opsclaw-cli jobs list
opsclaw-cli scan project::env::target
opsclaw-cli postmortem <incident-id>
opsclaw-cli completions zsh

Run opsclaw-cli --help or opsclaw-cli <command> --help for the complete surface. Commands label themselves [local], [remote], or [local|remote].

Build and test

OpsClaw requires Rust 1.88 or newer.

cargo build --release --locked
cargo test
cargo clippy --all-targets --all-features -- -D warnings

The two binaries are target/release/opsclaw and target/release/opsclaw-cli. A bare-binary deployment must provide its own reachable Hindsight service through HINDSIGHT_URL.

The repository parity suite tests every persistence contract against both the in-memory implementation and SQLite. Browser journeys live in web-tests/; live model evaluations are opt-in through just evals because they spend tokens.

Further reading

• Architecture and invariants
• Sandboxed evaluation setup
• Operator-facing changes
• Project brief