harness-agent 0.6.0

Multi-provider coding agent CLI + SDK

Harness

State-of-the-art open-source coding agent

CLI + SDK that works with any LLM — Claude, GPT, Gemini, Ollama, or any OpenAI-compatible endpoint.

The only open-source agent to score 100% on Harness-Bench and outperform Claude Code, OpenCode, and pi-mono.

Quick Start · Screenshots · Commands · Providers · Features · SDK · Benchmarks · Tutorials · Contributing

---

⚡ Quick Start

Install

curl -fsSL https://raw.githubusercontent.com/AgentBoardTT/openharness/main/install.sh | bash

Or with pip:

pip install harness-agent

Connect

harness connect

Pick a provider, paste your API key, done. Your key is saved to ~/.harness/config.toml.

Where do I get an API key?

• Anthropic (Claude): https://console.anthropic.com/settings/keys
• OpenAI (GPT): https://platform.openai.com/api-keys
• Google (Gemini): https://aistudio.google.com/apikey

Use

# Interactive REPL
harness

# One-shot command
harness "Fix the authentication bug in auth.py"

# Bypass mode — full auto-approve, no prompts (great for CI/scripts)
harness --permission bypass "Run all tests and fix failures"

# Use a specific model
harness -p openai -m gpt-5.2 "Refactor this function"

# Use a local model (no API key, fully private)
harness -p ollama -m llama3.3 "Write unit tests for utils.py"

# Resume a previous session
harness --session abc123 "Continue where we left off"

That's it. You're running a state-of-the-art coding agent.

back to top

---

📸 Screenshots

REPL Banner

Interactive Command Palette

Type / to open a filterable command palette. Arrow keys navigate, Enter selects, Escape dismisses.

Agent Execution

Status & Models

back to top

---

🎛 Interactive REPL Commands

Type / in the REPL to open the command palette, or use any command directly.

Command	Description
/help	Show available commands and tips
/connect	Set up or change your API key
/model	Switch model (e.g. /model gpt-5.2)
/models	List available models
/plan	Plan implementation with a read-only agent
/review	Review code changes or a specific file
/team	Decompose a task and run agents in parallel
/status	Show provider, model, session, and cost
/cost	Show token usage and cost for this session
/compact	Summarize conversation to free up context
/session	Show or switch session ID
/diff	Show git diff of changes in working directory
/init	Create a HARNESS.md project config file
/doctor	Check your setup (provider, API key, tools)
/permission	View or change permission mode
/clear	Clear the screen

back to top

---

🔌 Providers

Harness works with every major AI provider — switch with a single flag.

Provider	Models	How to connect
Anthropic	Claude Opus 4.6, Sonnet 4.6, Haiku 4.5	harness connect and choose Anthropic
OpenAI	GPT-5.2, GPT-4.1, o3, o4-mini, GPT-4o	harness connect and choose OpenAI
Google	Gemini 2.5 Pro, 2.5 Flash, 2.0 Flash	harness connect and choose Google
Ollama	Llama, Mistral, Qwen, Phi, etc.	No key needed — runs locally
OpenAI-compatible	DeepSeek, Groq, OpenRouter	--base-url flag

harness models list          # Browse 50+ supported models
harness models info sonnet   # Get details for a specific model

back to top

---

🛠 Features

Built-in Tools

Tool	What it does
Read	Read file contents
Write	Create or overwrite files
Edit	Find-and-replace inside files
Bash	Run shell commands
Glob	Find files by name pattern
Grep	Search inside files with regex
Task	Spawn sub-agents for parallel work
WebFetch	Pull content from web pages
AskUser	Ask you a question mid-task
Checkpoint	Save/restore file snapshots

Sub-Agents

The agent can spin up specialized workers in parallel:

Agent	Access	Use Case
general	Full tools	Complex multi-step tasks
explore	Read-only	Fast codebase exploration
plan	Read-only	Architecture planning
review	Read-only	Structured code review

Permission Modes

You control what the agent can do:

Mode	Behavior
default	Reads are automatic, writes ask for approval
accept_edits	File edits are automatic, shell commands ask
plan	Read-only — nothing gets changed
bypass	Full auto-approve (for scripts/CI)

Interactive Command Palette

Type / in the REPL to open a filterable dropdown. Arrow keys navigate, typing filters, Enter selects, Escape dismisses. All 16 slash commands are accessible instantly.

Async Steering (Live Message Injection)

While the agent is executing, type a message and press Enter to inject it between turns. The steering channel queues your message and the agent processes it at the next turn boundary — no need to wait for it to finish.

Context Compaction

When the conversation approaches the model's context limit (85% threshold), Harness automatically summarizes earlier messages while preserving key information. The compacted context targets 50% of the window, giving you room to keep working without starting a new session.

MCP (Model Context Protocol)

Connect external tool servers — Jira, Slack, databases, anything with an MCP adapter:

async for msg in harness.run(
    "Search our Jira board",
    mcp_servers={
        "jira": {
            "command": "npx",
            "args": ["-y", "@anthropic/mcp-server-jira"],
            "env": {"JIRA_TOKEN": "..."},
        }
    },
):
    ...

Skills

Teach the agent custom workflows by dropping a .md file in .harness/skills/:

---
name: deploy
description: Deploy to production
user_invocable: true
---

1. Run the test suite: `pytest tests/ -v`
2. Build the Docker image: `docker build -t myapp .`
3. Push to registry and deploy

Hooks

Run your own commands before/after every tool call:

hooks = [
    harness.Hook(
        event=harness.HookEvent.PRE_TOOL_USE,
        command="echo 'About to run {tool_name}'",
        matcher="Bash",
    ),
]

async for msg in harness.run("Fix the tests", hooks=hooks):
    ...

Memory

• Project instructions — Drop a HARNESS.md in your project root
• Auto-memory — Learnings persist across sessions in ~/.harness/memory/

back to top

---

🐍 SDK

Use Harness as a Python library to build your own tools on top of it.

Basic Usage

import harness

async for msg in harness.run("Fix the bug in auth.py"):
    match msg:
        case harness.TextMessage(text=t, is_partial=False):
            print(t)
        case harness.ToolUse(name=name):
            print(f"Using tool: {name}")
        case harness.Result(text=t, total_tokens=tok):
            print(f"Done ({tok} tokens): {t}")

With Configuration

async for msg in harness.run(
    "Refactor the database module",
    provider="openai",
    model="gpt-4.1",
    permission_mode="accept_edits",
    max_turns=50,
):
    ...

Sub-Agent API

from harness.agents.manager import AgentManager

mgr = AgentManager(provider=provider, tools=tools, cwd=".")
result = await mgr.spawn("explore", "Find all API endpoints")

# Parallel execution
results = await mgr.spawn_parallel([
    ("explore", "Find all API endpoints"),
    ("explore", "Find all database models"),
    ("review", "Review the auth module"),
])

Steering Channel (Async Message Injection)

Inject messages into the agent loop while it's running:

from harness.core.steering import SteeringChannel

steering = SteeringChannel()

# Start the agent with steering enabled
async for msg in harness.run(
    "Refactor the API layer",
    steering=steering,
):
    print(msg)

# From another coroutine, inject a message between turns:
await steering.send("Actually, skip the auth endpoints")

back to top

---

📊 Benchmark Results

Harness was benchmarked against the leading coding agents on 8 real-world tasks covering multi-file editing, bug fixing, error recovery, refactoring, context understanding, and code analysis.

Overall Scores

Agent	Claude Opus 4.6	GPT-5.2
Harness	7/8 (88%)	8/8 (100%)
Claude Code	7/8 (88%)	—
OpenCode	7/8 (88%)	7/8 (88%)
pi-mono	7/8 (88%)	8/8 (100%)

Harness is the only open-source agent that achieves a perfect score — and it does so across providers, not locked to one.

Per-Task Breakdown (GPT-5.2)

Task	Harness	OpenCode	pi-mono
Multi-file editing	PASS (17.5s)	PASS (19.4s)	PASS (26.8s)
Error recovery	PASS (5.2s)	PASS (11.7s)	PASS (10.1s)
Tool efficiency	PASS (1.8s)	PASS (5.6s)	PASS (9.2s)
Context understanding	PASS (9.7s)	FAIL	PASS (41.3s)
Project creation	PASS (3.0s)	PASS (7.6s)	PASS (3.8s)
Bug fixing	PASS (5.5s)	PASS (12.9s)	PASS (10.0s)
Code analysis	PASS (1.9s)	PASS (5.2s)	PASS (2.3s)
Refactoring	PASS (6.4s)	PASS (11.7s)	PASS (12.7s)

Speed

Agent	Model	Avg per Task	Total (8 tasks)
Harness	GPT-5.2	6.4s	51.0s
Harness	Opus 4.6	12.5s	99.7s
Claude Code	Opus 4.6	16.4s	131.5s
OpenCode	GPT-5.2	10.7s	85.8s
pi-mono	GPT-5.2	14.5s	116.2s

Harness is 2x faster than the next-fastest agent on GPT-5.2, and 30% faster than Claude Code on Opus.

back to top

---

⚙️ Configuration

Config File

Created automatically by harness connect. Lives at ~/.harness/config.toml:

[providers.anthropic]
api_key = "sk-ant-..."

[providers.openai]
api_key = "sk-..."

Environment Variables

If you prefer env vars, those work too:

export ANTHROPIC_API_KEY="sk-ant-..."
export OPENAI_API_KEY="sk-..."
export GOOGLE_API_KEY="AIza..."

back to top

---

🏗 Architecture

src/harness/
  core/
    engine.py          Top-level run() entry point
    loop.py            Agent loop (provider -> tools -> repeat)
    session.py         JSONL session persistence
    context.py         Context window management + compaction
    config.py          Config loading (env, TOML, HARNESS.md)
    steering.py        Async steering channel for live message injection
  providers/
    anthropic.py       Claude adapter
    openai.py          GPT / OpenAI-compatible adapter
    google.py          Gemini adapter
    ollama.py          Ollama local model adapter
    registry.py        Model catalogue (50+ models)
  tools/               Read, Write, Edit, Bash, Glob, Grep, Task, Web, etc.
  agents/              Sub-agent registry + lifecycle manager
  hooks/               Pre/post tool-use hook system
  mcp/                 MCP client + progressive tool discovery
  skills/              Skill loader (SKILL.md parser)
  memory/              Auto-memory + project instructions
  permissions/         Permission rules engine
  ui/                  Rich terminal output + streaming + diffs
  eval/                SWE-bench, Harness-Bench, metrics, reports
  cli/                 Click CLI entry point + subcommands

back to top

---

🧪 Evaluation

Run Benchmarks

# Quick validation — 8 tasks, ~$1
harness eval harness-bench --provider anthropic --model sonnet

# SWE-bench Lite — 300 real GitHub issues
harness eval swe-bench --split lite --max-tasks 10

# List benchmarks
harness eval list

Available Benchmarks

Benchmark	Tasks	Description
Harness-Bench	8	Multi-file editing, error recovery, refactoring, analysis
SWE-bench Lite	300	Curated subset of real GitHub issues
SWE-bench Verified	500	Human-verified solvable issues
SWE-bench Full	2,294	Complete benchmark

back to top

---

🔧 Development

git clone https://github.com/AgentBoardTT/openharness.git
cd openharness
uv pip install -e ".[dev]"
uv run pytest tests/ -v
uv run ruff check src/ tests/

---

📚 Tutorials

Learn Harness from zero to enterprise with our step-by-step tutorial series:

#	Tutorial	Difficulty	Time
1	Getting Started — Your First AI Coding Agent	Beginner	10 min
2	The Python SDK — Building AI-Powered Tools	Beginner	15 min
3	Permission Modes & Safety Controls	Intermediate	15 min
4	Budget Controls & Cost Optimization	Intermediate	15 min
5	Policy-as-Code — Governing Agent Behavior	Intermediate	20 min
6	Audit Logging & Compliance	Intermediate	20 min
7	Sandboxed Execution	Intermediate	20 min
8	Sub-Agents & Parallel Execution	Intermediate	20 min
9	CI/CD Integration — AI Agent in Your Pipeline	Advanced	25 min
10	Enterprise Production Deployment	Advanced	30 min

👉 Browse all tutorials

back to top

---

🤝 Contributing

We'd love your help. Here's how:

• Bug reports — Open an issue
• Feature requests — Open an issue
• Pull requests — Fork, branch, submit

Areas where we especially need help:

• New provider adapters
• Additional tools
• Benchmark tasks and evaluation
• Documentation and examples

---

📄 License

MIT

The best agent scaffold is an open one.