harnessctl

One command for Claude Code, Codex, DeepAgents, Gemini, Cursor, OpenCode, and custom coding agents.

harnessctl is a small CLI that lets you run, compare, hand off, and fail over between coding-agent CLIs without changing how you work.

Use it when one agent is better at planning, another is better at editing, and a third is available when the first one hits a rate limit.

Install

npm install -g harnessctl
harnessctl setup

Or use Homebrew:

brew install mnvsk97/tap/harnessctl

Other options:

# Linux / macOS binary installer
curl -fsSL https://raw.githubusercontent.com/mnvsk97/harnessctl/main/install/install.sh | bash

# From source
git clone https://github.com/mnvsk97/harnessctl.git
cd harnessctl
bun install
bun run src/cli.ts --help

Quick Start

# Run the default agent
harnessctl run "fix the auth bug"

# Pick an agent
harnessctl run --agent codex "refactor the database layer"
harnessctl run --agent deepagents "run the tests and fix failures"

# Open an interactive agent shell
harnessctl shell --agent claude

# Check configured agents
harnessctl doctor

Core Workflows

Hand Off Between Agents

Every run gets a run ID. Use it to continue with another agent.

harnessctl run --agent codex "refactor auth"
# run: 1713364500000-codex

harnessctl handoff 1713364500000-codex --agent claude "review and add tests"

The next agent gets a compact handoff: original task, summary, changed files, and a pointer to the full context file under .harnessctl/handoffs/.

Compare Agents

Run the same prompt across agents and optionally ask a judge agent to pick the best result.

harnessctl compare "fix this bug" --agents codex,claude
harnessctl compare "fix this bug" --agents codex,claude --judge claude

Run Pipelines

Chain agents through different stages of a task.

harnessctl pipeline "build auth module" --plan codex --build claude --test codex

Fail Over Automatically

Configure a fallback agent for rate limits, token limits, or auth failures.

# ~/.harnessctl/agents/claude.yaml
fallback: codex
auto_failover: true
failover_transfer: transcript

When Claude fails for a limit/auth reason, harnessctl hands the task to Codex with context.

Useful Commands

harnessctl list                         # installed/configured agents
harnessctl doctor                       # health checks
harnessctl models --agent codex         # known models for an agent

harnessctl logs                         # recent run history
harnessctl logs --run-id RUN_ID         # one run only
harnessctl replay RUN_ID                # rerun a previous prompt
harnessctl stats --cost                 # spend summary

harnessctl run "continue" --resume      # resume latest harness session
harnessctl run --name auth-fix "fix"    # name a session
harnessctl run --template review "src"  # use a prompt template
harnessctl run --budget 2.00 "task"     # daily spend guardrail

harnessctl context set "Node 22, postgres, follow existing patterns"
harnessctl context sync

Supported Agents

Agent	Native resume	Transcript handoff	Failover
Claude Code	yes	full	full
Codex	no	full	full
DeepAgents	yes	summary + session DB pointer	summary
Gemini	yes	full	summary
Cursor	yes	full	summary
OpenCode	no	no	summary
Custom YAML	configurable	no	summary

Custom agents live in ~/.harnessctl/agents/<name>.yaml.

Setup Checklists

Each underlying coding harness must be installed and authenticated before harnessctl can run it. Use harnessctl doctor after setup to confirm the selected harness is ready.

Claude Code setup

Required before use:

claude --version
claude auth login
harnessctl doctor

Harnessctl invokes Claude Code in print mode and reads native session JSONL files from ~/.claude/projects for richer handoffs.

Codex setup

Required before use:

codex --version
codex login
harnessctl doctor

Codex can authenticate with ChatGPT login or an API key. Harnessctl invokes Codex with codex exec for headless runs.

DeepAgents setup

DeepAgents support uses the CLI's non-interactive stdin mode, so it behaves like the other headless harnessctl adapters:

harnessctl run --agent deepagents "fix the failing tests"
harnessctl run --agent deepagents "continue that change" --resume
harnessctl run --agent deepagents -- --agent backend-dev

The adapter invokes deepagents --stdin --auto-approve --shell-allow-list recommended --quiet --no-stream. To pass DeepAgents-specific flags, put them after -- or in ~/.harnessctl/agents/deepagents.yaml under extra_args. Project context syncs to DeepAgents' native .deepagents/AGENTS.md file.

Before running DeepAgents through harnessctl, configure provider credentials with DeepAgents' /auth flow, or set OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, GOOGLE_CLOUD_PROJECT, a custom api_key_env from ~/.deepagents/config.toml, or the matching DEEPAGENTS_CLI_* variant in your shell, project .env, ~/.deepagents/.env, ~/.harnessctl/.env, project .harnessctl/.env, or ~/.harnessctl/agents/deepagents.yaml under env.

OpenAI-compatible gateways can be configured directly in DeepAgents, then used by harnessctl without passing model flags every time:

# ~/.deepagents/config.toml
[models]
default = "openai:provider-account/model-name"

[models.providers.openai]
base_url = "https://gateway.example.com"
api_key_env = "GATEWAY_API_KEY"
models = ["provider-account/model-name"]

[models.providers.openai.params]
use_responses_api = false

# ~/.deepagents/.env
GATEWAY_API_KEY=...

Check setup:

deepagents --version
deepagents -n "Say hello" -q --no-stream
harnessctl doctor

Gemini setup

Required before use:

gemini --version
harnessctl doctor

Authenticate by running gemini interactively, or set one of the supported environment paths before running harnessctl:

export GEMINI_API_KEY=...
# or use Vertex AI configuration
export GOOGLE_GENAI_USE_VERTEXAI=true

Harnessctl invokes Gemini with stream JSON output and --yolo for headless execution.

Cursor setup

Required before use:

agent --version
agent login
harnessctl doctor

You can also set CURSOR_API_KEY instead of using agent login. Harnessctl invokes Cursor's agent CLI in print/headless mode.

OpenCode setup

Required before use:

opencode --version
opencode auth login
harnessctl doctor

Harnessctl invokes OpenCode with opencode --pipe, so the OpenCode CLI must already be configured with whatever provider credentials your OpenCode setup needs.

Custom YAML harness setup

Custom harnesses live at:

~/.harnessctl/agents/<name>.yaml

Minimal shape:

cmd: your-agent-command
args: ["--headless"]
stdin: true
timeout: 300
env:
  YOUR_AGENT_API_KEY: ${YOUR_AGENT_API_KEY}

Then verify:

harnessctl list
harnessctl doctor
harnessctl run --agent <name> "Say hello"

Configuration

First run creates:

~/.harnessctl/
  config.yaml
  agents/
  runs/
  sessions/
  templates/
  pipelines/

Common settings:

harnessctl config set default codex
harnessctl config set-fallback claude codex --auto
harnessctl config get

Development

bun install
bun run src/cli.ts --help
bun test
bun run typecheck
bun run bundle
bash test/sim-headless-failover.sh
bash test/sim-fallback.sh
bash test/sim-deepagents.sh
bash test/real-deepagents-smoke.sh

test/real-deepagents-smoke.sh uses configured DeepAgents provider credentials when available. If no credential is configured, it runs the installed DeepAgents CLI with a temporary local class_path chat model so the harnessctl subprocess path is still exercised end to end without external secrets.

Docs

More detailed guides live in docs/guide, and example notebooks live in docs/examples/notebooks.

License

MIT