Open-source AI coding tools fall into three groups: assistants that work inside the editor, agents that run in the terminal, and autonomous agents that plan and carry out multi-step tasks with little supervision. The right choice depends on where your work happens, how much you want to review each step, and how you plan to control cost and repeatability. This comparison uses the catalog data for eight projects to show how they differ.

Criteria for comparing options

Before picking a tool, check these points:

  • Interface: Does it run in the terminal, as an IDE-integrated agent, or as an MCP server that other agents call?
  • Level of autonomy: Does it wait for approval at each step, or does it execute a whole task on its own?
  • Cost visibility: Can you see how many tokens and how much money each task uses?
  • Repeatability: Can you define a workflow once and run it the same way again?
  • Extensibility: Does it support skills, plugins or external tool servers?
  • License: Most options here are Apache-2.0 or MIT, which allows commercial use and modification. Check the license field for each project before adoption.

Terminal agents

qwen-code

qwen-code is an open-source AI coding agent that runs in your terminal. It is licensed under Apache-2.0 and has about 28,500 stars. Its main strength is that it fits into an existing shell workflow without requiring a specific editor. The trade-off is that you work through the terminal, so developers who prefer a graphical IDE may need an adjustment period.

DeepSeek-Reasonix

DeepSeek-Reasonix is described as a coding agent for complex software engineering tasks. It is MIT-licensed and has about 35,800 stars. It targets multi-file work where a simple completion tool falls short. The catalog description does not list its supported models or integrations, so verify those in the repository before relying on them.

rtk

rtk is a CLI proxy rather than an agent. It sits in front of common development commands and reduces LLM token consumption by 60 to 90 percent, according to its description. It is a single Rust binary with no dependencies and is Apache-2.0 licensed with about 83,500 stars. It does not replace an agent; it lowers the cost of running one. Pair it with a terminal agent if token spend is your main concern.

Editor-integrated agents

oh-my-pi

oh-my-pi is a coding agent with the IDE wired in, built by Stencil Labs. It is MIT-licensed with about 35,600 stars and is listed in the AI Tools category. Its strength is the tight link between the agent and the editor. The trade-off is that tying the workflow to an IDE can make it harder to run the same agent headless or in automation.

Archon

Archon describes itself as an open-source harness builder for AI coding, aimed at making AI coding deterministic and repeatable. It is MIT-licensed with about 23,700 stars. Choose it when consistency across runs matters more than improvised exploration.

Autonomous agents and tool servers

agentic-awesome-skills

agentic-awesome-skills is a local, agent-first control plane for catalog discovery, selection, stack validation and planning. It is backed by more than 2,600 agentic skills and includes a CLI, a local MCP server, plugins and a Workbench. It is MIT-licensed with about 47,700 stars. Its value is organizing which skills an agent should use, so it suits teams that manage many skills rather than a single assistant.

serena

serena is an MCP toolkit for coding that provides semantic retrieval and editing. It is listed under MCP Servers with about 30,300 stars; the license field is not provided in the catalog data, so check it before use. Because it is an MCP server, any MCP-compatible agent can call it, which makes it a building block rather than a standalone assistant. More MCP options are in the MCP servers catalog.

codeburn

codeburn is a free, local tool that tracks AI coding token usage and cost across 37 tools and agents, including Claude Code, Cursor, Codex and Gemini, broken down by model, project and task. It is MIT-licensed with about 11,500 stars. It does not write code; it tells you what your agents cost. Use it alongside any of the agents above. The developer tools section lists related utilities.

Who should pick what

  • Terminal-first teams should start with qwen-code and add rtk to cut token use on routine commands.
  • IDE-centric developers should evaluate oh-my-pi, keeping in mind the trade-off around automation.
  • Teams that need repeatable output should look at Archon for structured workflows.
  • Teams managing many skills or tools should consider agentic-awesome-skills and serena as shared infrastructure.
  • Anyone worried about spend should install codeburn first, whatever agent they choose, so cost decisions rest on measured data.

Recommendation

Do not choose on star counts alone. Start with the interface you already use, decide how much autonomy you accept, and measure cost from the first week. A practical setup is one terminal or editor agent, a token-reduction proxy such as rtk, and a cost tracker such as codeburn. Add MCP servers and skill catalogs once you know which capabilities your team actually needs. Browse the Developer Tools and AI Tools categories to compare more options.