harness

Run any coding agent from a single CLI. Harness spawns the selected agent as a subprocess, translates its native streaming output into a unified NDJSON event stream, and outputs it to stdout.

Write one integration, and it works with any supported agent backend.

Supported agents

Install

Recommended — curl installer:

curl -fsSL https://harness.lol/install.sh | sh

Via crates.io:

cargo install harnesscli

Build from source:

git clone https://github.com/ayshptk/harness-cli.git
cd harness
cargo build --release
# Binary at target/release/harness

Quick start

# Run Claude Code with a prompt
harness run --agent claude --prompt "explain this codebase"

# Use a model alias
harness run --agent claude --model sonnet --prompt "fix the bug"

# Dry-run — see the resolved command without executing
harness run --agent claude --model sonnet --prompt "hello" --dry-run

# Pipe prompt from stdin
echo "explain this codebase" | harness run --agent claude

# List available agents
harness list

# Check if an agent is installed
harness check claude --capabilities

Model registry

Harness maps human-friendly model names to the exact IDs each agent expects:

# List all known models
harness models list

# Resolve an alias for a specific agent
harness models resolve opus --agent claude
# → claude-opus-4-6

# Update the registry cache
harness models update

Built-in aliases include opus. You can add your own in harness.toml, and the cached registry at ~/.harness/models.toml is auto-updated from GitHub.

Configuration

Place a harness.toml in your project root (or any parent directory):

default_agent = "claude"
default_model = "sonnet"
default_permissions = "full-access"
default_timeout_secs = 300

[agents.claude]
model = "opus"
extra_args = ["--verbose"]

[models.my-model]
description = "My custom model"
provider = "anthropic"
claude = "my-custom-model-id"

Unified event stream

Every agent's output is translated into a common NDJSON format with 8 event types:

• SessionStart — session initialized
• TextDelta — streaming text chunk
• Message — complete message
• ToolStart — tool invocation beginning
• ToolEnd — tool invocation complete
• UsageDelta — incremental token usage and cost update
• Result — run finished
• Error — error occurred

Documentation

Full docs at harness.lol

License

MIT