A terminal coding agent that reads your codebase, plans changes, edits files, runs commands, and verifies its work.

python-agent-harness is inspired by gptel-agent-harness and opencode. It brings opencode's prompts and core behaviors—such as AGENTS.md discovery, plan/build modes, skills, sub-agents, and todo tracking—into a lightweight Python implementation with only three runtime dependencies:

• rich
• httpx
• prompt_toolkit

It works with any OpenAI-compatible API and is designed to be easy to inspect, customize, and use for everyday software development.

Demo

Quick start

Install from PyPI

pip install python-agent-harness

python-agent-harness config --init
python-agent-harness run

Optional extras:

pip install "python-agent-harness[mcp]"   # MCP server integration
pip install "python-agent-harness[dev]"   # development tools

Install from source (GitHub)

git clone git@github.com:beacoder/python-agent-harness.git
cd python-agent-harness

make install
. venv/bin/activate

python-agent-harness config --init
python-agent-harness run

Optional extras:

pip install -e ".[mcp]"   # MCP server integration
pip install -e ".[dev]"   # development tools

Edit ~/.config/python-agent-harness/config.json and set your base_url, api_key, and model.

Features

• FSM-driven execution — explicit WAIT / TOOL / TRET / SUPERVISE / DONE / ERRS / ABRT states. Completion supervision nudges the model when it stops early, while failed tool calls are sanitized so they never strand the agent. Transient API failures (429 / 5xx) retry with exponential backoff and jitter. Auth-expired status codes (configurable via AUTH_REFRESH_STATUS_CODES in config.py; defaults to [401, 502]) trigger automatic API key re-read from config/env — some API gateways return 502 instead of 401 when the backend auth token has expired. Note: codes in this list are treated as auth-expired exclusively and will not be retried with backoff, so only include codes that are unambiguously auth-related in your environment.
• Context management — CJK-aware token estimation, per-model context windows, and automatic compaction at 70% usage.
• Coding tools — Agent, TodoWrite, Glob, Grep, Read, Insert, Edit (including unified diffs), Write, Mkdir, Bash, Skill, Question, and PlanExit. Synchronous tools execute sequentially, but a round made up entirely of read-only tools (Read, Glob, Grep, Skill) is dispatched concurrently via a bounded thread pool; asynchronous tools such as Bash and Agent can run concurrently as well. Results are always delivered in the model's emitted order.
• Plan / Build modes — plan mode is read-only except for the per-session plan file.
• Persistent sessions — sessions are automatically saved after every response to ~/.local/share/python-agent-harness/sessions/, with LLM-generated titles and support for /restore --latest and /sessions.
• Focused TUI — a Rich-based interface with a pinned status bar, Todos panel, inline red/green diff rendering for Edit and Write, and a prompt_toolkit editor with history and completion. Esc+Enter submits, Ctrl-D quits, and Ctrl-C cancels without leaving the application.
• MCP support — optional MCP integration through the [mcp] extra. MCP tools become ordinary agent tools such as mcp__<server>__<tool>. Supports stdio, streamable-http, and sse transports.
• Slash commands — built-in /init, /review, /explain, and other commands, plus custom commands loaded from prompts/commands/*.md.
• Custom agents — switch the main agent's system prompt at runtime with /agent. Agent prompt files live in prompts/agents/*.md. Use default_agent in the config file to start sessions with a specific agent.

Inspired by opencode

Most of opencode's prompts and core behaviors have been ported to this project. The goal is to retain its practical coding-agent workflow while keeping the implementation small, dependency-light, and easy to customize.

Prompt and behavior mapping

The following opencode prompts have corresponding implementations in python-agent-harness:

Configuration

All LLM settings live in a single JSON configuration file. Environment variables are optional.

{
  "llm": {
    "base_url": "https://api.openai.com/v1",
    "api_key": "sk-...",
    "model": "gpt-5-mini",
    "reasoning_effort": null,
    "stream": true
  },
  "models": {
    "_comment": "Named LLM profiles for /model switching. Partial settings; unset keys inherit the main llm.",
    "deepseek": {
      "base_url": "https://api.deepseek.com/v1",
      "model": "deepseek-chat"
    },
    "qwen": {
      "base_url": "https://dashscope.aliyuncs.com/compatible-mode/v1",
      "model": "qwen3.5-coder"
    }
  },
  "context_windows": {
    "_comment": "Optional per-model context-window overrides (tokens). Keys are model names or substrings (e.g. deepseek-v4 = 1000000); matched in file order, first match wins. Overrides the built-in CONTEXT_WINDOWS table in config.py.",
    "deepseek-v4": 1000000
  },
  "subagent_llm": {
    "profile": null,
    "base_url": null,
    "api_key": null,
    "model": null,
    "temperature": null,
    "max_tokens": null,
    "timeout": null,
    "reasoning_effort": null,
    "stream": null
  },
  "default_agent": null,
  "paths": {
    "context_path": null,
    "skill_path": null
  },
  "mcp": {
    "servers": {
      "example": {
        "transport": "stdio",
        "command": "npx",
        "args": [
          "-y",
          "@modelcontextprotocol/server-filesystem",
          "/tmp"
        ],
        "env": [],
        "parallel": false,
        "timeout": null,
        "enabled": false
      }
    }
  }
}

Configuration options

• llm — main LLM configuration. Optional keys include temperature, max_tokens, timeout, reasoning_effort, and stream. Values such as reasoning_effort are passed to the API as-is when set. run --no-stream overrides stream.
• models — named LLM profiles for runtime switching with /model. A profile is a partial settings dictionary; unset keys inherit from the main llm. default restores the main LLM configuration.
• context_windows — optional per-model context-window overrides (tokens). Keys are model names or substrings (e.g., deepseek-v4); matched in file order, first match wins. Overrides the built-in CONTEXT_WINDOWS table in config.py.
• subagent_llm — LLM configuration for Agent tool requests. Unset values inherit from the main llm. Set profile to reuse a profile from models. Precedence is: profile settings > explicit subagent_llm settings > main llm > environment variables.
• default_agent — name of the agent to use at session start (instead of the built-in agent.md). The agent must exist as a .md file in the prompts/agents/ directory. When unset or null, the built-in default agent is used. Use /agent default in the TUI to switch back to the built-in at any time.
• paths.context_path / paths.skill_path — locations from which to load context files and skills. When unset, the project-local <project>/contexts and <project>/skills directories are used.
• mcp.servers — MCP server configuration. Requires the [mcp] extra. Each server supports transport, command, args, env, url, headers, parallel, timeout, and enabled.
• Configuration precedence — code defaults < config file < OPENAI_* environment variables. Sub-agent settings also support OPENAI_SUBAGENT_* (_BASE_URL, _API_KEY, _MODEL).
• Custom config — use --config PATH or PYTHON_AGENT_HARNESS_CONFIG.
• LLM logging — request and response bodies are logged as JSON to /tmp/python-agent-harness-<date>-<id>.json. Set LLM_LOG_DIR to change the directory. The log path is printed at startup.

Usage

python-agent-harness run [project-dir]

Launches the interactive TUI agent. If project-dir is omitted, the current directory is used.

Slash commands

Custom commands from prompts/commands/*.md are registered as slash commands as well (TUI only).

Custom agents

Agent prompt files are markdown files (.md) placed in the prompts/agents/ directory. Each file becomes a switchable agent profile available via the /agent TUI command.

Commands vs Agents

Project layout

python_agent_harness/
├── agent.py           # Agent FSM core: states, transitions, supervision
├── tool_runner.py     # Tool-call execution/delivery + history salvage
├── context_manager.py # Context-ratio tracking + compaction
├── client.py          # OpenAI-compatible streaming client (httpx)
├── models.py          # Message / ToolCall / ToolSpec data classes
├── token_estimator.py # CJK-aware token estimation + calibration
├── planmode.py        # Plan/build modes + plan-file lifecycle
├── prompts.py         # Prompt loading + system-prompt assembly
├── persistence.py     # Session persistence + titles
├── session.py         # Session wiring hub + MCP lifecycle
├── subagent.py        # Sub-agent runner + error containment
├── commands.py        # Init/review/custom command definitions
├── cli.py             # CLI entry points
├── tui/               # Rich + prompt_toolkit TUI (package)
├── diffrender.py      # Unified diff generation + Rich rendering
├── mcp/               # Optional MCP client
└── tools/             # Tool implementations + registry

Development

Requires Python ≥ 3.11. CI runs against Python 3.11, 3.12, and 3.13 on Linux, MacOS and Windows.

make test                           # unit tests
venv/bin/pip install -e ".[dev]"    # development tools
venv/bin/ruff check .               # lint
venv/bin/pyright                    # type checking
venv/bin/python -m build            # build sdist + wheel
venv/bin/pip-audit                  # dependency audit

CI blocks on Ruff and Pyright failures.

Design philosophy

Keep it intact, not bloated.

The project aims to provide a capable coding-agent within a lightweight framework.

Related projects

• gptel-agent-harness — the Emacs-based implementation that inspired this project.
• opencode — the primary source of many prompts and coding-agent behaviors.

License

MIT