Incident investigation
Move from alert to evidence.
Trace crash loops, degraded rollouts, failed jobs, noisy logs, and cloud symptoms across the systems already in your path.
Autonomous operations.Bounded authority.
Investigate infrastructure, diagnose incidents, and execute tightly scoped remediation through fail-closed policy, least-privilege credentials, budget stop-losses, and verifiable audit chains.
Default denyUnknown actions stop.
No shell surfaceStructured argv only.
Chain of recordEvery decision linked.
YOU
Why is api-0 crash-looping in web?
prod · read only
01Loaded prod read-only credential
02kubectl describe pod api-0
03kubectl logs api-0 --previous
04Correlated exit 137 with heap config
Root cause isolated
Container OOMKilled. JVM heap is 22% above the workload memory limit.
$0.0841
audit chain verifiedsha256 · 8 events
Connected systems
KubernetesGitHubAWSGoogle CloudAzureSSH
one safety coreOperational range
Useful across the incident. Restrained at the boundary.
The agent gets enough context to form a real diagnosis. Authority is granted separately, for a specific action, target, environment, and amount of time.
CI recovery
Explain the failure. Propose the fix.
Inspect GitHub Actions runs, connect the failing job to source, and keep remediation inside a reviewable pull request.
Cloud posture
Read broadly. Write narrowly.
Investigate AWS, Google Cloud, and Azure with curated command families while provider-wide deployment and IAM changes stay denied.
Guarded operations
Turn intent into a bounded change.
Apply, roll out, scale, or roll back only when policy, credential scope, dry-run, approval, and an expiring capability all agree.
Execution model
The prompt is never the permission.
A model can ask. It cannot grant itself a credential, widen a target allowlist, bypass a policy deny, disable a dry run, or approve its own production change.
01
Request
Natural-language intent enters through an identity-aware interface.
02
Budget
Call, tool, recursion, time, and USD stop-losses are checked.
03
Policy
Unknown tools, targets, flags, and configuration fail closed.
04
Credential
One scoped credential family is selected for the winning rule.
05
Execute
A structured argv request runs with shell expansion disabled.
06
Audit
Decision, execution, cost, and resolution events join one hash chain.
Fail-closed by construction
Missing configuration, unpriced models, unknown tools, unmatched commands, and unavailable credentials refuse to boot or deny the action.
Read the policy pipelineStop-losses at every layer
Per-run calls, tool invocations, recursion, wall time, and USD cost meet per-principal daily budgets and hard executor timeouts.
Understand budgetsCredentials are architecture
Remote executors split credentials by environment and read/write channel, then accept only signed decisions that match their fixed identity.
Inspect the boundaryOne agent, many entry points
Meet operators where incidents already happen.
Every interface enters the same gateway, budget, policy, credential, approval, and audit path. There is no “convenient” side door.
CLI
Stream investigations from a local REPL with per-turn cost and policy feedback.
Operations UI
Private agent chat, live runs, approvals, policy decisions, spend, and audit integrity.
Slack
Thread-aware chat-ops with identity mapping and the same escalation boundary.
HTTP + webhooks
Receive Alertmanager and GitHub events through authenticated, deduplicated routes.
Scheduler
Run drift, certificate, backup, and hygiene checks with fixed anti-overlap controls.
Remote executor
Keep write credentials inside isolated environment-and-channel workers.
Architecture
One graph. Explicit control planes.
The reasoning loop is separated from execution, policy, credentials, state, and audit. That keeps integrations composable without turning the model into a trusted control plane.
Explore the architectureentry
Interfaces
Events
reason
Agent graph
Budgets
Policy
act
Executor
Credential
hash-chained audit
Start locally
Your first read-only investigation takes one workspace.
Install the released tool, initialize a workspace, provision a secrets-denied Kubernetes identity, then choose the contexts the agent may see.
uv tool install "opendevops[checkpoint,ssh]==0.1.2"- 1Initialize the configuration workspace
- 2Add a scoped model and target credential
- 3Run
opendevops config check