Warp Factories infrastructure and security

Warp Factories separates execution, inference, and credential boundaries. You choose where code runs, which providers serve inference, and which credentials each agent receives. Warp coordinates the work and stores transcripts and artifacts.

Control plane and execution plane

Each factory uses two planes:

• Control plane - Warp coordinates runs, identity and configuration, observability, integrations, storage, and inference routing.
• Execution plane - A Warp-hosted sandbox or a managed self-hosted worker checks out code, runs setup, invokes tools, builds the project, and executes commands.

flowchart LR
  I["Integrations and triggers"] --> C["Warp control plane<br/>coordination · identity/config<br/>observability · inference routing"]
  C --> H["Warp-hosted sandbox"]
  C -->|"task, config, and scoped<br/>runtime credentials"| S["Managed self-hosted worker"]
  H -->|"results, transcripts,<br/>artifacts, telemetry"| C
  S -->|"results, transcripts, attachments,<br/>artifacts, and telemetry<br/>can contain code context"| C
  C --> P["Warp-managed or<br/>customer-configured inference"]
  C --> D["Warp-managed storage"]

With managed self-hosting, repository checkouts, commands, and sandbox files stay on machines you control. Prompts, results, transcripts, attachments, artifacts, and telemetry still flow through Warp and configured providers. See deployment patterns, execution security, and the data security and boundaries reference for details.

Runners

A runner selects the operating system, architecture, sandbox image, and vCPU and memory shape. The factory’s definition supplies repositories, setup commands, and secrets. Define runners in runners/*.yaml; agentDefaults.runner sets the default, and agents and automations can override it. The execution host determines how each setting is used:

Setting	Warp-hosted	Direct	Kubernetes	Docker
platform.os, platform.arch, platform.mac.version	Selects the sandbox OS, architecture, and macOS VM version	Does not configure the host; the worker host determines the platform	Use Linux; the cluster and pod_template schedule nodes	Use Linux; the Docker daemon determines the architecture
platform.linux.dockerImage	Linux sandbox image	Ignored	Task and setup container image	Task container image
platform.linux.registryCredentialSecretName	Warp-managed private-registry credential	Ignored	Ignored; use imagePullSecrets	Ignored; configure registry credentials on the worker, such as in its Docker config
instanceShape	Sandbox size, subject to plan limits; omitted uses the workspace default	Ignored; the task uses host resources	Task-container requests and limits; overrides matching pod_template resources; omitted keeps template or cluster defaults	Task-container limits; omitted adds no limits

Self-hosted runner settings don’t provision capacity. Worker configuration controls resources, registry authentication, volumes, placement, and backend policy. See runner compute options for available configurations.

Choose an execution host

A factory runs its work on one of two execution hosts: Warp-hosted compute or a worker in the managed self-hosting architecture.

Decision area	Warp-hosted	Managed self-hosted
Compute	Warp provisions the sandbox	Your team provisions the worker
Checkout and commands	Run on Warp-managed compute	Run on your infrastructure
Network	Warp manages sandbox connectivity	The worker connects outbound to Warp; no inbound firewall port
Private services	Must be reachable from the hosted sandbox	Reachable through the worker’s network access
Operations	Warp manages capacity and lifecycle	Your team manages capacity, isolation, updates, and availability

To route factory work to a managed self-hosted worker (an Enterprise feature):

1. Deploy a worker - Review the self-hosting requirements, then connect a worker that authenticates to Warp with a self-hosted worker API key. Workers run on linux/amd64 and linux/arm64, and the worker’s platform determines which workloads it can run.
2. Pair it with a compatible runner - Choose a runner that matches the worker’s platform.
3. Select the worker in the factory definition - Set workerHost so the factory routes work to it.

For a working definition with workerHost set and a platform-matched runner, see 07-self-hosted-worker.

Unmanaged self-hosted agents and other CLI agents can’t serve as a factory’s execution host, but they can exchange work with a factory through Factory MCP.

Managed self-hosting structures

The worker backend sets isolation and scheduling on your infrastructure:

Structure	How factory runs execute	What your team operates	Use it when
Docker (default)	In a separate Docker container on the worker host	The worker daemon, host, Docker daemon, images, capacity, and container policy	Docker is available and you want per-run container isolation without Kubernetes
Kubernetes	As a Kubernetes Job in the worker’s namespace	The worker deployment, cluster, namespace RBAC, scheduling, admission policy, and capacity	Your team already operates Kubernetes or needs cluster-native policy and scheduling
Direct	In a separate workspace directly on the worker host, sharing its OS and kernel	The worker daemon, host security, dependencies, capacity, and cleanup	A container runtime isn’t available or runs need direct access to host resources

Choose inference independently

Execution hosting doesn’t select inference. Configure that boundary separately.

Inference options for factory runs

Factory agents execute as cloud agents, so only inference options that support cloud agent runs apply:

Option	Support for factory runs	Team controls	Key boundary
Warp-managed inference	Supported	The model selected for each agent	Warp provides the provider account and bills model usage with Warp credits
Team-managed keys and endpoints	Supported on Business and Enterprise	Shared OpenAI, Anthropic, or Google keys, or an OpenAI-compatible endpoint	Warp stores the encrypted credential and uses it only at the inference boundary; it never enters the worker or run environment
BYOLLM: AWS Bedrock	Supported on Enterprise	The AWS account, IAM role, available Claude models, and provider billing	Warp assumes your IAM role through OIDC; inference runs in your AWS account
BYOLLM: Gemini Enterprise	Not supported for factory runs	Interactive inference in your Google Cloud project	Gemini Enterprise BYOLLM currently supports interactive agent requests only

Self-serve BYOK and custom inference endpoints are stored on an individual member’s device and don’t apply to factory runs. With team-managed keys and endpoints, select a specific provider model or endpoint; Auto continues to use Warp-managed inference. When you supply the provider, its retention follows your account and contract, and Warp can’t enforce ZDR for that provider.

Credential boundaries

A factory handles four kinds of credentials, each with its own boundary:

Credential	Used for	Boundary
Inference credentials	Model provider requests	Used only at the inference boundary; never injected into the sandbox
Execution secrets	APIs, package registries, and tools an agent uses	Delivered from an explicit per-agent allowlist; factory agents that don’t act as a specific user receive no managed secrets by default
Harness authentication	Third-party harnesses such as Claude Code or Codex	Configured separately from the agent’s secret allowlist
Repository identity	Checking out code and pushing changes	Runs act with the creating user’s authorization (changes are attributed to them) or as the agent itself for unattended work; set by the definition’s credentialStrategy

Scope each credential to the resources and actions its agent needs. Warp redacts known secret values at output boundaries, but redaction is a backstop, not a substitute for narrow external permissions and rotation. See cloud agent secrets, harness authentication, secret redaction, and team identity for the underlying controls.

Factory roles and permissions

Factory creation uses the team-level Warp Factories role. Team and workspace admins and team-level Editors can create factories.

Three roles govern an existing factory:

• User - View and run the factory, and propose definition changes.
• Editor - Edit the factory definition, configure its agents, repositories, and automations, attach existing integrations, secrets, and MCP servers, and select existing runners.
• Admin - Manage members, delete the factory, and force-merge a blocked definition change.

Team and workspace admins and owners always have the Admin role. Admins for a factory manage its roles; team and workspace admins manage the team-level role.

Managing shared runners and connecting team-wide integrations and providers require team or workspace permission. Factory roles don’t control specification or pull request review. Treat factory definition changes as operational code, and protect merges through repository settings.

Metering

Warp meters hosted compute, Warp-provided inference, and platform services. Self-hosting moves compute costs to your infrastructure; customer-supplied inference bills your provider account. Platform services still consume credits. See platform credits for details.

Deployment checklist

1. Classify access - Inventory repositories, data, and services.
2. Set boundaries - Choose execution and inference hosts.
3. Scope credentials - Configure allowlists, harness authentication, and repository identity.
4. Enforce review - Protect factory-definition and pull request merges.
5. Validate operations - Test egress, isolation, rotation, capacity, observability, and metering.

Related pages

• Deployment patterns - Compare execution and inference boundaries.
• Runners - Configure compute and worker compatibility.
• Managed self-hosting - Deploy and operate a worker backend.
• Execution security - Review data boundaries, egress, and backend controls.
• Team-managed model keys and endpoints - Configure inference credentials.
• Roles and permissions - Govern team and workspace resources.
• Platform credits - Understand factory metering.