Pre-integrated agents

Run agents supported by Harbor.

Harbor includes integrations for many agents so that you don’t have to install any dependencies.

CLI

export OPENAI_API_KEY="..."
harbor run \
  -t hello-world/hello-world \
  -a codex -m openai/gpt-5.6-sol

Config

config.json

{
  "agents": [
    {
      "name": "codex",
      "model_name": "openai/gpt-5.6-sol"
    }
  ],
  "tasks": [
    {
      "name": "hello-world/hello-world"
    }
  ]
}

export OPENAI_API_KEY="..."
harbor run --config config.json

Python

import asyncio

from harbor.job import Job
from harbor.models.job.config import JobConfig
from harbor.models.trial.config import AgentConfig, TaskConfig


async def main():
    job = await Job.create(
        JobConfig(
            agents=[
                AgentConfig(
                    name="codex",
                    model_name="openai/gpt-5.6-sol",
                )
            ],
            tasks=[TaskConfig(name="hello-world/hello-world")],
        )
    )
    await job.run()


asyncio.run(main())

Use -a to select the agent and -m to select its model. Run harbor run --help for the current options.

Available agents

Agent names

aider, antigravity-cli, antigravity-sdk, claude-code, cline-cli, codex, computer-1, copilot-cli, cortex-code, cursor-cli, deerflow, devin, dspy-rlm, eve, fx, gemini-cli, goose, grok-build, hermes, junie, kimi-code, kimi-cli, langgraph, mcode, mimo, mini-swe-agent, nemo-agent, openclaw, opencode, openhands, openhands-sdk, oracle, nop, pi, qwen-coder, rovodev-cli, swe-agent, terminus-2, trae-agent, vibe

You can also run agents from the ACP registry or add a custom agent.

Agent capabilities

Capabilities describe optional features supported by each integration.

Capability	Purpose	Supported agents
ATIF	Produce an ATIF trajectory	acp, antigravity-cli, antigravity-sdk, claude-code, cline-cli, codex, computer-1, copilot-cli, cortex-code, cursor-cli, devin, eve, gemini-cli, goose, grok-build, hermes, junie, kimi-cli, mimo, mini-swe-agent, nemo-agent, openclaw, opencode, openhands, openhands-sdk, qwen-coder, rovodev-cli, swe-agent, terminus-2, trae-agent, vibe
Resume	Continue the native session between task steps	aider, claude-code, codex, copilot-cli, cortex-code, gemini-cli, goose, junie, kimi-code, kimi-cli, mcode, mimo, opencode, pi, qwen-coder
Load native trajectory	Restore an agent’s native session	claude-code, codex
Load ATIF trajectory	Load prior ATIF context	claude-code, codex
Handoff	Continue a completed session locally	claude-code
Native configuration	Accept the agent’s native configuration format	claude-code, codex, deerflow, junie
Windows	Run in Windows task containers	nop, oracle
ACP bridge	Act as the target in a simulated-user trial	claude-code, gemini-cli

Agent options

--ak and --ae are two of the most commonly used agent flags. Run harbor run --help to see the full set.

Flag	Purpose
--ak, --agent-kwarg	Pass a setting to the agent implementation. Values such as numbers, booleans, lists, and objects are parsed into their corresponding types.
--ae, --agent-env	Set an environment variable during agent setup and execution. Values remain strings.

Each agent supports its own kwargs and environment variables. Check the agent’s integration before setting them.

harbor run \
  -t hello-world/hello-world \
  -a claude-code -m anthropic/claude-sonnet-5 \
  --ak max_turns=20 \
  --ae MCP_TIMEOUT=20000

Here, max_turns=20 configures the Claude Code integration, while MCP_TIMEOUT is available to the agent process. Both flags can be repeated.

Native configuration

Load native config file for corresponding agent. Check out Claude Code settings.json and Codex configuration config.toml as example. claude-code, codex, deerflow, and junie accept a native config file:

# Config file
harbor run ... -a codex --ak "config=path/to/config.toml"

Inline JSON is also supported. Harbor converts inline JSON to the agent’s native config file format and passes it to the agent:

# Inline config
harbor run ... -a codex --ak 'config={"model_context_window": 200000}'