Built for Google Cloud's All Things Agentic HackathonLearn more →
Solvan
Get started
Open source · Autonomous production engineering

The open-source production engineer that never sleeps.

Solvan detects production degradation from real alert events, investigates with read-only evidence tools, acts within bounded authority, and verifies recovery — then stays on the case until the defect is permanently repaired.

Get startedSee how it worksStar on GitHub
scripts/bootstrap && scripts/start

Open source under Apache 2.0 — built for production systems on Google Cloud: Cloud Run, GKE, and Cloud SQL. Every action authorized, bounded, and receipted.

Observes and actuates across your GCP production environment

Your production environment

  • Google Cloud
  • Cloud Run
  • GKE
  • Cloud SQL
  • Cloud Monitoring
  • Cloud Logging
  • PostgreSQL

Your incident and tooling layer

  • PagerDuty
  • Slack
  • GitHub
  • Prometheus
  • Grafana
  • OpenTelemetry
  • Terraform

Platform

Built for Google Cloud production

Solvan deploys into your GCP project and region. Production credentials never reach a model.

01

Observe

Telemetry, audit history, and GKE events through typed read-only tools — every evidence record carries source, time range, and hash.

02

Actuate

A private actuator in your region holds the only production permissions. Bounded mitigations run under policy; high-risk actions need digest-bound approval.

03

Govern

Every agent runs under its own least-privilege identity. Policy screens every call, and all state stays in Cloud SQL in your project.

GCP is the first supported environment. AWS and Azure are on the roadmap.

How it works

From alert to verified recovery

One continuous loop — no hand-offs between tools, no lost context.

  1. 01

    Detect

    Events, not prompts.

    The evaluator runs every 25 seconds against real alert events. Duplicates are deduplicated; each incident opens with durable typed state.

  2. 02

    Investigate

    Read-only by construction.

    Agents gather evidence in parallel with typed read-only tools. Every record carries provenance — a hypothesis never silently becomes root cause.

  3. 03

    Act

    Bounded authority, exact approvals.

    Pre-authorized mitigations execute immediately. High-risk actions need approval bound to the exact action digest, target, and expiry.

  4. 04

    Verify & own

    Mitigation is not repair.

    Recovery is judged independently from fresh telemetry and probes. A Reliability Case stays open until permanent repair is verified.

Why Solvan is different

Built like an operations system, not a chatbot

Authority is structural

Permissions live in identity, IAM, and gateway policy — not prompt promises. A policy-denied action stays denied, even if a human approves it.

Blocked by policy

Mitigation ≠ repair

An incident can read “Service restored · repair open” for days while the Reliability Case drives the permanent fix.

Service restored · repair open

Evidence remains evidence

Confidence never converts a hypothesis into root cause. Every committed fact carries its source and freshness.

Cloud Monitoring · fresh

Durable typed state

Incidents, approvals, and receipts live in Cloud SQL — not a chat window. Agents can crash mid-incident and resume without duplicating side effects.

Checkpoint committed

Capabilities

Everything an on-call engineer is accountable for

Reliability Cases

Multi-day ownership with durable checkpoints. A case closes only when permanent repair is independently verified.

Exact approvals

Every high-risk approval binds to the action digest, target version, and expiry. Change any of them and it's void.

Independent verification

Recovery is judged by service-owned profiles and synthetic probes — the planner can't weaken its own success criteria.

Gated memory

Only verified learnings with full provenance reach the Memory Bank. Never a guess, a fluke, or an unverified fix.

Fleet governance

Identity, gateway enforcement, and Model Armor apply uniformly to every agent — observable activity, auditable authority.

Operator console

Decisions and evidence without private chain-of-thought. Every claim links to its source, every action to its receipt.

Security

Autonomy you can audit

An agent that touches production has to earn authority structurally.

  • Least-privilege agent identities

    Each agent runs under its own scoped identity — no shared credentials, no ambient authority.

  • Zero secrets in prompts, traces, or the browser

    Credentials never enter model context, traces, or the DOM. Evidence is redacted at the tool boundary.

  • Policy outranks approval

    The gateway can deny a critical action even when a human approves it.

  • Regional data residency

    State, telemetry, and memory stay in your GCP region, in your project.

  • Full audit trail with receipts

    Every detection, approval, and execution produces a durable, queryable receipt in Cloud SQL.

  • Tenant isolation

    Incidents, cases, and memory are partitioned per tenant at the data layer.

  • Open source and self-hosted

    Apache 2.0. Solvan deploys into your own GCP project — every line is yours to audit.

Solvan is an early-stage project. These are architectural commitments enforced in code — not certifications.

Put Solvan on call

Detect the next degradation before your users report it — and get to a verified recovery, with receipts.

Get startedRead the docs