Wardby guardian robot carrying a budget shield
Wardby for engineering teams

Build an agent workforce. Keep control of the work.

Agent runners help a model complete a task. Wardby governs the work around it. Use one self-hosted control plane to decide whether an agent may run, what it may access, what it may spend, and what reviewable outcome it may produce.

OpenAI, Anthropic, or Bedrock Keep your infrastructure Budgets enforced before spend Explicit action authority One owned agent catalog
Run it locally Run it on GKE View on GitHub

Project status: the control plane, scheduler, budget groups, MCP server, native agents, isolated Codex and Claude Code workers, GitHub draft-PR flow, and local observability stack are implemented and tested. Production readiness and cloud deployment coverage are still being expanded.

01 / The platform path

One MCP control plane from idea to governed execution

Your team can design and operate agents conversationally, trigger them from delivery systems, or run them on a schedule. Every path passes through the same Wardby controls before a model, tool, repository, or dollar is touched.

Ask or trigger1
CC
Claude Code
Build, configure, run, and inspect agents through MCP tools.
CX
Codex
Use the same governed catalog from a different coding workbench.
CI
Delivery events
GitHub Actions, or any CI that can call a webhook, triggers approved agents on PR, comment, merge, or manual dispatch.
24h
Schedules & signals
Recurring reviews, webhooks, metrics, incidents, and on-demand operations.
Govern and orchestrate2
Wardby control plane
Self-hosted
The durable operating boundary between intent and agent execution.
Agent catalogOwned definitions, prompts, tools, models, and schedules
Budget admissionReserve spend before a request reaches the provider
Policy & identityScoped repositories, tools, secrets, datastores, networks, and principals
Triggers & workflowMCP, CI events, webhooks, schedules, and retries
Execution routingSelect the approved builder, model, and isolated runtime
Evidence & metricsRun history, actual cost, metrics, and outcomes
Nothing runs until identity, policy, and available budget agree.
Build or execute3
CX
Codex builder
Works in an isolated checkout and returns a reviewable patch or pull request.
codetestsPR
CC
Claude builder
Uses the same bounded job contract with a separately reviewed driver.
codetestsPR
API
Native agent runtime
Runs analysis, reviews, monitoring, reporting, and tool workflows without changing code.
toolsreportsactions
Your infrastructure and boundaries
Docker / VM / on-prem GKE reference AWS portable runtime Other container platforms PostgreSQL
●
The whole system: conversations, delivery events, and schedules all use one catalog and one set of guardrails. Teams can change the model or builder without changing the governance contract.

02 / The reusable harness

Six agent systems. One operational contract.

Your framework becomes a managed suite rather than six unrelated scripts. Each agent can use Codex, Claude, or a native model flow while inheriting the same budgets, permissions, evidence, and operator-defined review policy.

Full-cycle builder

Delivery Pipeline

Turns an approved request into a plan, implementation, tests, review, and a draft pull request.

AnalyzeDesignBuildTestReview
System reviewer

Architecture Review

Examines the codebase for architecture, privacy, accessibility, reuse, coverage, and structural risk.

MapScoreExplainRecommend
Maintenance builder

Security Maintenance

Finds vulnerable components, separates safe fixes from breaking changes, verifies, and proposes a patch.

ScanTriagePatchVerify
Quality builder

QA End to End

Maps customer journeys, ranks coverage gaps by risk, and incrementally builds the missing test suite.

DiscoverRankWriteRun
Operational agent

System Monitoring

Turns production signals into a clear defect, evidence package, escalation, and approved follow-up work.

ObserveCorrelateReportEscalate
Portfolio agent

Project Tracking

Pulls actuals, budget, schedule, ownership, and progress into one live view with early drift detection.

CollectCompareForecastAlert

03 / Centralized budget intelligence

Know who spent it, why, and for which feature

Shipped today: hard per-run limits and shared daily, weekly, or monthly budget groups that stop additional model work at the cap. Next, with a work-tracking connector, a feature key can travel with the request into Wardby. Every child agent and model call inherits that context, so reserved budget and final agent expense can roll up to the owning feature, epic, team, and portfolio. The example below shows the target Jira pattern.

Agent expense by Jira feature and owner
Jira-linked integration pattern
Jira cardFeature / ownerBudgetAgent activityActual
PAY-241
Payment retry recovery
MCBilling Platform · Maya C.
$1,20062% consumed
PlannerCodex buildClaude reviewQA
$74418 runs
MOB-118
Receipt capture refresh
LOMobile · Liam O.
$85034% consumed
ArchitectureClaude buildE2E QA
$2899 runs
SEC-092
Dependency hardening
ARSecurity Engineering · Ava R.
$50081% consumed
ScannerCodex patchVerifier
$40527 runs
1. Work is selectedA connector supplies the work key, feature, team, owner, and approved limit.
2. Budget is reservedWardby admits only work that fits the feature and portfolio limits.
3. Children inherit contextEvery builder, reviewer, retry, and model call remains attached to the parent feature.
4. Actuals roll upFinal usage appears by run, feature, owner, team, and portfolio.

Scope: this view attributes Wardby-managed agent and model expense. Labor, infrastructure, or other delivery costs appear only when imported from their source systems. Jira integration is shown as the target connector pattern, not as a claim that the connector is currently shipped.

04 / What the team receives

Autonomy with a receipt

Wardby does not replace engineering judgment. It makes delegated work repeatable, bounded, and easy to inspect, whether the result is code or an operational decision.

PR
Optional reviewable code changesPlan, patch, tests, evidence, and a draft pull request when the agent is authorized for code
Review gate
R
Architecture and security reportsPrioritized findings with context and remediation options
Traceable
!
Operational defect packagesSignal, diagnosis, impact, evidence, and proposed next action
Escalated
$
Budget and portfolio actualsReserved spend, final cost, run history, and program drift
Measured

05 / Boundaries you can verify

Self-hosted, provider-neutral, and honest about scope

Wardby does not replace your agent frameworks, gateways, or observability tools. It is the operating boundary around them, and it runs in infrastructure you control.

Security boundaries

  • Sandboxed tools: native tools run in constrained QuickJS with controlled fetch and secrets.
  • Isolated coding workers: hardened containers with resource limits and reviewed egress.
  • Protected remote MCP: HTTP is an OAuth 2.1 resource server; stdio trusts the local operator.
  • Sanitized operations: events and metrics exclude prompts, code, diffs, and credentials.
Security boundaries →

Host it in your cloud

  • Local Docker, VM, or on-prem: portable PostgreSQL and container workflow.
  • Production baseline: separate runtime and migration images with a Compose/Caddy boundary.
  • GKE Autopilot: supported reference with private Cloud SQL and isolated worker pods.
  • AWS and other clouds: portable runtime and Bedrock adapter; native AWS module planned.
Deployment targets →

Bring your own observability

  • Prometheus metrics on a private /metrics endpoint, pulled by your collector.
  • A local Prometheus and Grafana profile with cost, budget, and run dashboards.
  • Scrape into AWS CloudWatch, Google Cloud Monitoring, or any Prometheus-compatible platform.
Observability →

Where Wardby fits

  • Repository automation, agent frameworks, and coding agents run the work.
  • Gateways and provider dashboards see model calls; observability tools trace them.
  • Wardby adds owned agents, admission-time budgets, scoped capabilities, and run history across them.
  • Budget views cover Wardby-managed work, not every AI expense in an organization.
Full comparison →

Run locally in five minutes

Start without cloning Wardby

Bring Node.js 24, Docker, and an OpenAI or Anthropic API key. Quickstart launches PostgreSQL, creates a $1 demo agent, and can connect Claude Code or Codex through MCP.

npx --yes @wardby/cli@latest quickstart
Local guide GKE guide Identity provider guide
For developers

Use Claude Code or Codex to call capable agents without memorizing deployment, identity, and budget mechanics.

For engineering leaders

Scale repeatable agent workflows across teams while retaining cost, risk, and outcome visibility.

For the platform team

Run one self-hosted control plane across approved models, builders, CI systems, and cloud boundaries.