v1.0 loading...

YOLO Mode. 10x'd.

Stop typing. Start shipping.

Agent Harbor is the vibe engineering harness built for long-horizon tasks. Orchestrate dozens of Claude, Gemini and Codex agents, with an advanced local sandbox and effortless rollbacks.

A sandbox built for YOLO mode

  • Near-instant agent startup and perfect environment replication in a copy-on-write file system
  • Roll back to any local snapshot and fork agents to continue prompting with your agent/model of choice
sandbox-cow — zsh

➜ah

Starting copy-on-write environment...

Replicating local configs... Done (12ms)

Snapshot 0d8F2A created.

Ready for autonomous execution.

➜ah task --agent claude --prompt "Create a new file 'hello.py' that prints 'Hello, Agent Harbor!'"

Anti-halting features for long-horizon tasks

  • Automated diagnosis of stuck processes to overcome deadlocks, infinite loops and connection failures in test suites
  • Supervisor agent conducts targeted web research to overcome challenges and re-prompt for higher quality responses
agent-harbor
Settings
New Task
Describe what you want the agent to do...|
agent-harbormainclaude-3-5-sonnet (x1)
⏎ Go
Existing tasksRepo [All]Status [All]Creator [All]
Implement payment processing
ecommerce-platform•feature/payments•claude-3-5 (x1)•5 min ago
Evaluating performance implications
Implementing proper error handling and logging
Optimize database queries
analytics-platform•main•gpt-4 (x1)•10 min ago
Considering edge cases...
Enter Launch draftShift+Enter Insert newlineCtrl+? Shortcut help

Designed for professional development

  • Let agents run test suites and validate patches on all operating systems in parallel with advanced leader/follower orchestration
  • Every task, command, and transcript can be recorded for later review, debugging, and compliance
validation-suite
Config
Active Suite
Validating pull request#1249on all targets
agent-harborfix/race-condition
● Live
Validate Patch #1249 (MacOS)
macOS-14•test-runner-mac
Running
Running integration tests...
Verifying FS snapshots
Validate Patch #1249 (Linux)
ubuntu-22.04•test-runner-linux
Running
Building Docker container...
Installing dependencies
Running 2 jobsQueued 0 jobs

PRIVATE

Agent Harbor does not collect any data from model usage and you bring your own subscription, so data never enters Agent Harbor's servers

SECURE

Customize the sandbox at file system and network level to granularly control what agents have access to. Fully auditable workflow.

FLEXIBLE

Agent Harbor integrates with a wide variety of terminals and IDEs, supports remote task management through web and mobile interfaces and allows teams to set up on-prem execution clusters.

OPEN SOURCE

Fully open sourced and extensible. See our guidelines if you would like to learn how to contribute.

Join the Early Access Program

If you are an avid vibecoder and would like to be part of a selective community of early testers, you're in the right place.

Get a direct line to the Agent Harbor team for support, see under the hood before public launch, score exclusive swag, and help us build the best product for you.