ashcode 0.1.0

A terminal coding agent in Python — streaming agentic loop, 12 built-in tools, MCP servers, subagents, and a configurable approval layer.

AshCode

A terminal coding agent in Python — streaming agentic loop, 12 built-in tools, MCP servers, subagents, context compaction, and a configurable approval layer for shell and file operations.

AshCode reads and writes files, runs shell commands, searches the web, and delegates work to scoped subagents — driven by any OpenAI-compatible model endpoint.

Install

pip install ashcode

Requires Python 3.11+.

Configure

AshCode talks to any OpenAI-compatible endpoint — OpenRouter, OpenAI, Together, or a local llama.cpp or vLLM server. Create a .env in the directory you want to work in:

API_KEY=your-openrouter-api-key-here
BASE_URL=https://openrouter.ai/api/v1

The model must support tool calling.

Use

ashcode                                      # interactive session
ashcode "explain the retry logic in this repo"   # single prompt
ashcode --cwd ../other-project               # point it somewhere else

Inside an interactive session, /help lists the slash commands — switching model or approval policy mid-session, inspecting token usage, saving and resuming sessions, and creating checkpoints.

What it does

Agent loop — fully streaming, parallel tool calls within a turn, a typed event stream that decouples the loop from any renderer, and a configurable turn budget.

12 built-in tools — read_file write_file edit_file apply_patch with unified-diff previews, list_dir glob grep for navigation, shell with timeout and env scrubbing, web_search web_fetch, and memory todo for cross-turn state.

Context management — automatic compaction into a structured summary at 80% of the window, tool-output pruning outside a protected recent window, and per-turn token accounting.

Safety — six approval policies from on-request through yolo, regex classification of shell commands, mandatory confirmation for writes outside the working directory, and secrets scrubbed from the subprocess environment.

Extensibility — MCP servers over stdio and HTTP/SSE, custom tools auto-discovered from .ai-agent/tools/*.py, lifecycle hooks around runs and tool calls, and AGENT.MD picked up as project instructions.

Configuration is TOML, merged from a user-level file and a project-level .ai-agent/config.toml. Secrets stay in .env and never enter the config file.

Writing a custom tool

Drop a Python file in .ai-agent/tools/ and subclass Tool. Note the import path is ashcode.tools.base when working against the installed package:

from ashcode.tools.base import Tool

Tools are self-describing — declare a Pydantic schema and it is converted to an OpenAI function schema automatically. No change to the agent loop or the prompt is needed.

Links

• Source, full documentation and engineering notes: github.com/AuthRan/AshCode
• Video demos: reading through a codebase · creating and deleting a file
• Issues: github.com/AuthRan/AshCode/issues

MIT licensed.