Warp Oz

Product Overview

Warp Oz is the cloud agent orchestration platform from Warp, the company behind the popular AI-native terminal.^[1] While most coding agent tools run locally, Oz is cloud-first — agents execute on Warp's infrastructure (or yours) with unified visibility across desktop, web, and mobile. As of June 2026 the product is still branded Oz (the "Oz Agent Platform") and is in active development.^[1]

Oz positions itself as "the orchestration platform for cloud agents" — turning agents into automations that run without your machine being active.^[1] It launched February 10, 2026; Warp says it's used by over 700,000 developers at companies including Docker, Ramp, and Peloton, plus over half of the Fortune 500.^[2]

What It Does

Core capabilities:^[3]

• Schedule agents like cron jobs (automations)
• Trigger agents from Slack mentions, Linear comments, or webhooks
• Multi-repo changes in a single task
• Unified control plane across desktop, web, mobile, CLI
• Session sharing — teammates can join running agent sessions
• Full API/SDK for programmatic orchestration

Key primitives:^[3]

• Triggers — Events that start work (schedules, integrations, API calls)
• Tasks — Units of tracked work with inputs, state, and outputs
• Environments — Docker image + repos + setup commands for consistent execution
• Hosts — Where execution happens (Warp cloud or your infrastructure)

Multi-Harness Orchestration (May 2026)

In May 2026, Warp upgraded Oz into a multi-harness control plane — it now runs Claude Code and OpenAI Codex alongside Warp Agent:^[4]

• Single pane of glass — consistent access controls, governance policies, and audit logs across all harnesses
• Automatic subagent orchestration — Oz deploys and tracks multiple agents in parallel for long-horizon tasks (feature builds, migrations, deployments)
• Cross-harness Agent Memory — persistent memory that carries what works for your team across sessions, repos, and harnesses
• Harness routing — Warp positions Warp Agent for long-running coding, Claude Code for bug investigation and frontend work, Codex for migrations^[1]

This is a meaningful repositioning: Oz is no longer just a runner for Warp's own agent — it's pitching itself as the enterprise control plane for whichever coding agents your team already uses.^[5]

Integrations

Oz has first-party integrations with Slack and Linear:^[6]

When triggered, Oz:

1. Reads the conversation/issue context
2. Spins up your configured environment
3. Runs the workflow in the cloud
4. Posts progress and PRs back to the same conversation

This is a key differentiator — most local-first tools (Emdash, Conductor) don't have Slack/Linear integration.

Multi-Model Support

Oz supports a broad range of models:^[7]

OpenAI:

• GPT-5.2 (low/medium/high/extra-high reasoning)
• GPT-5.2 Codex, GPT-5.1 Codex Max, GPT-5.1, GPT-5

Anthropic:

• Claude Opus 4.6 (default and max)
• Claude Opus 4.5 and Sonnet 4.5 with thinking mode
• Claude Opus 4.1, Haiku 4.5, Sonnet 4

Google:

• Gemini 3 Pro, Gemini 2.5 Pro

Auto modes:

• Auto (Cost-efficient) — Optimizes for lower credit consumption
• Auto (Responsive) — Prioritizes speed with best available model
• Auto (Genius) — Selects most capable model for complex tasks

Environments

Environments ensure consistent execution across triggers:^[3]

# Create an environment with guided flow
/create-environment ./frontend ./backend

# Or specify repos
/create-environment https://github.com/your-org/api.git

An environment includes:

• Docker image — The toolchain and runtime
• Repositories — One or more repos to clone
• Setup commands — Dependency install, builds, bootstrapping

The same environment works across Slack, Linear, CLI, and API triggers — no per-trigger configuration duplication.

Self-Hosting (Enterprise)

For enterprises needing execution within their network, Oz supports self-hosted workers:^[8]

# Set API key
export WARP_API_KEY="your_team_api_key"

# Run the worker
docker run -v /var/run/docker.sock:/var/run/docker.sock \
  -e WARP_API_KEY="$WARP_API_KEY" \
  warpdotdev/oz-agent-worker --worker-id "my-worker"

# Route tasks to your worker
oz agent run-cloud --prompt "Refactor auth" --host "my-worker"

Self-hosting architecture:

• Oz orchestrator still manages task lifecycle and observability
• Execution happens on your infrastructure
• Code never leaves your network
• You get same session sharing and team visibility

The worker is open source: oz-agent-worker

CLI and SDK

Oz has a full programmatic interface:^[3]

# Run cloud agent from CLI
oz agent run-cloud --prompt "Fix the auth bug" --environment env-123

# Create integrations
oz integration create slack --environment env-123
oz integration create linear --environment env-123

SDKs:

• Python SDK
• TypeScript SDK

Use cases: CI pipelines, incident tooling, custom automation, internal dashboards.

Pricing

Warp offers credit-based pricing (as of June 2026):^[9]

BYOK (Bring Your Own Key) and custom inference endpoints are now available on every plan, including Free — use your own API keys for model providers.^[9]

Zero Data Retention

Warp has ZDR agreements with all model providers:^[7]

• LLM providers don't train on customer data
• Inputs/outputs deleted after generating responses
• Available across all plans by default

This is a significant enterprise feature — compliance teams care about model training data policies. Warp is also SOC 2 Type II certified.^[1]

Adoption

As of June 2026, concrete external Oz adoption numbers are thin. What Warp has published:

• Warp (the company) is used by 700,000+ developers, including Docker, Ramp, Peloton, and over half of the Fortune 500^[2]
• Internally, Warp says Oz writes 60% of its own PRs^[10]
• Backers include Sequoia Capital, GV, Sam Altman, Marc Benioff, and Dylan Field; total funding is $73M, with the $50M Series B led by Sequoia^[2][11]

Caution: the headline adoption figures are for Warp the terminal, not Oz the platform — and the "60% of our PRs" stat is self-reported with no external verification. Oz itself launched in February 2026 and shipped its biggest repositioning (multi-harness) in May 2026, so treat it as an early, fast-moving product.^[5]

Strengths

• Cloud-first — Agents run without your machine being active
• First-party integrations — Slack and Linear built-in
• Multi-model — Claude, GPT-5/Codex, Gemini with Auto selection
• Multi-harness — Orchestrates Claude Code, Codex, and Warp Agent with cross-harness memory
• Self-hosting — Keep execution in your network (Enterprise)
• Session sharing — Team visibility into running agents
• Full SDK — Python/TypeScript for programmatic control
• ZDR by default — Model providers don't train on your data
• Warp backing — Well-funded company ($73M from Sequoia, GV, and angels), active development
• Unified experience — Same agents across desktop, web, mobile, CLI

Weaknesses / Risks

• Warp lock-in (softening) — Historically tied to Warp terminal; multi-harness support reduces this, but the control plane is still Warp's
• Cloud dependency — Even self-hosted workers need Oz orchestration cloud
• Credit-based pricing — Costs can be unpredictable for heavy usage
• No Jira — Slack and Linear, but no Jira integration
• No signed commits — Missing compliance feature for regulated industries
• Newer product — Oz is newer than Warp terminal; less battle-tested
• macOS primary — Warp terminal is macOS/Linux; no Windows

What Developers Say

As of June 11, 2026, there is no substantive independent developer discussion of Oz to quote. The lone Hacker News submission of the Oz launch post drew zero comments,^[12] and in HN threads about Warp more broadly, the only people discussing Oz specifically are Warp's own founders. Warp the terminal has an active community; Oz the orchestration platform has not yet generated independent practitioner commentary. This section will be updated when real third-party sentiment emerges.

Competitive Landscape

Codex App (OpenAI) — Similar cloud execution + Skills/Automations, but OpenAI-only. Oz is multi-model — and as of May 2026 can orchestrate Codex itself as one of its harnesses.^[4]

Emdash — Local-first with issue tracker integration. No cloud execution, no Slack integration. Different architecture.

Conductor — Local worktrees for parallel Claude Code. No cloud, no integrations. Simpler tool.

GitHub Copilot Workspace — Cloud-based but GitHub-native. Less flexible than Oz for custom workflows.

Tembo — Agent-agnostic orchestration with Jira, signed commits, BYOK. Oz has Slack/Linear; Tembo has broader enterprise features.

Ideal User

Warp Oz is built for teams who:

• Already use Warp terminal (natural extension)
• Run Claude Code or Codex and want one control plane with shared governance and memory
• Want cloud execution without managing infrastructure
• Trigger work from Slack or Linear
• Need multi-model flexibility (Claude, GPT-5, Gemini)
• Value session sharing and team visibility
• Have enterprise needs (self-hosting, ZDR)
• Don't need Jira or signed commits

Bottom Line

Warp Oz is the cloud orchestration play from a well-funded terminal company. While most coding agent tools are local-first, Oz is cloud-first — agents run on Warp's infrastructure with Slack/Linear integrations built in.

The enterprise story is compelling: self-hosted workers keep code execution in your network while Oz handles orchestration and observability. The ZDR agreements with model providers address compliance concerns.

The May 2026 multi-harness update changed the calculus. Oz now orchestrates Claude Code and Codex alongside Warp Agent, with cross-harness memory and unified governance — it's no longer just a runner for Warp's own agent, it's a bid to be the control plane for everyone's agents.^[4] The lock-in critique has shifted from "Warp's agent only" to "Warp's cloud only." If you want agent-agnostic orchestration with broader enterprise integrations (Jira, signed commits), Tembo fills that gap.

Watch Oz carefully. Warp has significant resources, 700k+ developers on its terminal, and is shipping fast — Oz went from launch to multi-harness control plane in three months. As the "parallel agent orchestration" category matures, cloud-first vs. local-first will be a key architectural divide.

---

Research by Ry Walker Research • methodology

Disclosure: Author is CEO of Tembo, which competes in the agent orchestration space.