Open source tools for coding agents
Open source tools for coding agents from solo devs and small teams, ranked by GitHub stars on October 22, 2026: skills, token savers, memory, routing.
Most lists of open source AI tools are the same twenty company projects: LangChain, Ollama, vLLM, Dify, and so on. They are fine, but they are built by funded companies, and none of them need help paying for tokens. I wanted the other list. The tools people install next to Claude Code or Codex to make the agent cheaper, better behaved, or less forgetful, built by one person or a handful of them.
So here are twenty of the most starred of those, in six groups. The rules were simple: public source under an open source license, owned by a personal account or an org built around that one project, including a few that have since grown a company or a foundation around it, pushed to in the last 90 days, and useful while you work with a coding agent. No training frameworks, no model weights, nothing from a company you had heard of before the tool. I left out NextChat, which has turned into a hosted sign-up service. Star counts are from the GitHub API on October 22, 2026. Stars are not downloads, and I did not rank anything by how good I think it is. A few of these grew a paid tier after they took off, and I say so where it changes what you get.
Skills and methods
A skill is a Markdown file your agent reads before it works. These four change how the agent behaves on every task without you touching your prompts, and two of them have passed 280,000 stars.
Jesse Vincent's Superpowers is a whole development methodology packaged as skills. The agent brainstorms with you first, writes a design, plans it, then hands the plan to subagents that work test first. It installs from the Claude Code and Codex plugin marketplaces and also runs on Cursor, Gemini CLI, OpenCode, Pi, and about ten other harnesses. After OpenClaw it is the most starred project on this page.
Use it forA feature that takes more than one sitting, when you want the agent to agree on a design and write tests before it writes code.
Everything Claude Code, or ECC, by Affaan Mustafa is the maximalist option: 71 agents, 311 skills, hooks, memory, and a scanner that checks your agent config for security problems, in one plugin. It works best with Claude Code, installs natively into Codex, and has adapters for Cursor, OpenCode, and more. One maintainer ships weekly, funded by sponsors and a paid GitHub App for private repos.
Use it forOne install that covers planning, review, memory, and a security check across the two or three agents you switch between.
Addy Osmani's agent-skills is 25 skills behind nine slash commands that follow the lifecycle: spec, plan, build, test, review, ship. It is the tidy, opinionated end of this group, and the skills CLI installs it into 70-plus agents.
Use it forA light, predictable spec-to-ship flow when the full methodology feels like too much ceremony.
OpenSpec is spec-driven development for coding assistants. You propose a change, the agent writes the proposal, specs, design, and task list into a folder, works through the tasks, and archives the change when it is done. Plain Markdown, and built for codebases that already exist as much as for new ones.
Use it forAdding features to a codebase that already exists, when you want the spec and the task list to live in the repo next to the code.
Token austerity
This group did not exist a year ago. Agents read a lot and write a lot, every token has a bill attached, so people built tools that make them read and write less. Four of them have passed 75,000 stars.
Ponytail, from the DietrichGebert account, puts a lazy senior dev inside your agent. The skill makes it look at what already exists before writing anything, so a date picker becomes ten lines around the input component you already have, instead of a new library. The project's own benchmark on 39 tasks in Claude Code reports 53 percent less code and 45 percent fewer tokens with tests still written. It also ships a review command and a band in Claude Code that shows what the current task is costing.
Use it forAn agent that over-builds, pulling in libraries and writing helpers for things your repo already has.
Caveman by Julius Brussee started as a joke that worked: make the agent talk like a caveman and the same fix takes a third of the output tokens. Version 4 is six parts behind one installer: short answers, logs and test output compressed before the model reads them, a model and effort picker per request, and scripts the agent keeps and reuses. The README points to an Adobe Research paper and a JetBrains A/B test. The Auto picker needs a free Caveman account and sends your latest ask through their cloud; the other five parts stay local.
Use it forLong sessions where the explanations pile up. You keep the code and the fix, and lose the paragraphs around them.
RTK is a single Rust binary that sits between your agent and the shell. It filters and compresses the output of a hundred-plus common commands, git, test runners, package managers, before it lands in the context window, with savings the project puts at 60 to 90 percent on those commands. Read the license before you pick a download: the full edition mixes Apache 2.0 with BUSL 1.1 for the compression engine, and there is an Apache-only build called rtk-oss. That edition is the one that gets RTK on this list; the engine in the default download is source-available, not open source.
Use it forTest suites and git commands that dump hundreds of lines into the context every time the agent runs them.
Headroom compresses what the agent reads, tool outputs, logs, files, and RAG chunks, on your machine, and leaves the earlier conversation byte for byte alone so the provider cache still hits. Originals are cached so the model can ask for them back. It runs as a Python or TypeScript library, a proxy, an MCP server, or a one-command wrap around Claude Code, Codex, Aider, and a dozen other agents. There is a paid enterprise tier for teams.
Use it forBig JSON and log payloads in an agent you built yourself, or one command to wrap the CLI agent you already use.
Memory and context
Agents forget everything when the session ends and read your repo one file at a time. These three fix one or the other.
Graphify turns a codebase, with its docs, SQL schemas, and PDFs, into a knowledge graph your agent queries instead of grepping. Code is parsed with tree-sitter on your machine with no model involved, and every edge is labeled as extracted from the source or inferred. The free tool is Apache 2.0. The company behind it now sells an always-on cloud version.
Use it forDropping an agent into a large codebase it has never seen, so it can ask what calls what instead of reading files one at a time.
claude-mem by Alex Newman records what your agent did during a session, compresses it with a model, and injects the relevant bits into the next session. It began as a Claude Code plugin and now installs into OpenClaw, Codex, Gemini, OpenCode, pi, and others. The installer offers a hosted memory tier with a free trial, and the local observer is one flag away.
Use it forWork that runs over days, when the agent keeps relearning your project every morning.
Repomix by Kazuki Yamada packs a whole repository into one AI-friendly file, with the tree, the contents, and a token count, so you can hand a codebase to any model. It is the oldest tool in this group, there is a version that runs in the browser at repomix.com, and it shipped a release last month.
Use it forPasting a small repo into a chat model for a one-off review, or shipping a codebase snapshot to a model with no file access.
Agent frameworks and runtimes
The harness is the loop around the model: tools, state, and the terminal or chat surface it lives in. These three come from a GitHub org with about thirty public members, one with four, and one developer.
OpenClaw is the personal AI assistant you run on your own hardware. It shows up in Discord, Slack, Telegram, WhatsApp, iMessage, and a couple dozen other channels, and treats models and agent harnesses as plugins you can swap. Peter Steinberger started it in late 2025, it now sits under the OpenClaw Foundation, and it is the most starred repository on this page by a wide margin.
Use it forAn assistant on your own box that you reach from the chat app you already have open, with your keys and your data staying home.
pi is a minimal agent harness split into TypeScript packages: a unified LLM API, an agent loop, a terminal UI, and a coding agent CLI. It deliberately skips subagents and plan mode and expects you to add what you want through extensions and packages. OpenClaw is built on its packages.
Use it forBuilding your own agent or terminal tooling on a small base you can read end to end.
ruflo, formerly claude-flow, by rUv is a meta-harness that wraps Claude Code and Codex with swarms of specialized agents, shared memory, and a console that shows cost and events. The quick start is three Claude Code plugins; the full install is a CLI command, and any other client can reach it as an MCP server.
Use it forRunning several specialised agents on one job with shared memory, and a console that shows what it is costing.
Routing models
Three small tools that decide which model answers, so you are not tied to one provider or one laptop.
Claude Code Router gives Claude Code, Codex, OpenCode, pi, and a list of other agents one local endpoint, and you manage providers, keys, routing rules, and fallbacks behind it from a desktop app or the CLI. It logs every request with tokens, latency, and an estimated cost.
Use it forSwitching models or providers under your agent without editing its config, with a fallback when one provider is down.
Simon Willison's llm is a command line tool and Python library for talking to OpenAI, Anthropic, Gemini, and dozens of other models, remote or local, through plugins. Every prompt and response lands in SQLite, which makes it a good way to compare models before you point your agent at one.
Use it forTrying the same prompt against several models from the shell, with a log you can query later.
llama-swap by Benson Wong is a Go binary that loads and unloads local models on demand behind an OpenAI and Anthropic compatible endpoint, in front of llama.cpp, vLLM, and friends. One binary, one config file. It is the smallest project here by stars, and the README says thousands of people run it in front of their local models.
Use it forOne machine, several local models, swapped on demand behind a single endpoint your agent already speaks.
Chat interfaces
Chat windows over your own keys and local models: two desktop apps and one server you host. Cherry Studio and LibreChat run agents and MCP servers too.
Cherry Studio is an AI workstation for Windows, macOS, and Linux: chat with several models at once, agents with their own tool permissions, local knowledge bases, and MCP. It connects to cloud providers or to Ollama and LM Studio on your machine. AGPL 3.0, so read it before you fork.
Use it forA desktop workstation with chat, agents, and a knowledge base built from your own files.
LibreChat is the self-hosted ChatGPT clone that grew into a multi-user platform with agents, MCP, a sandboxed code interpreter, and an API for managing agents. It talks to Anthropic, OpenAI, Bedrock, Azure, Google, and any OpenAI-compatible endpoint.
Use it forA shared chat server for a team, with your own keys, user accounts, and agents everyone can reuse.
Chatbox is a desktop client for ChatGPT, Claude, and local models through Ollama, with your data stored on your device. The public repo is the community edition under GPL 3.0, kept in sync with a paid version.
Use it forA simple desktop chat over your own API keys and Ollama, with nothing to deploy.
Where boostmyagents fits
Most projects on this page are maintained by one person or a small team, and several are built with the very coding agents they improve. For many of those maintainers the token bill comes out of their own pocket. That bill is what boostmyagents exists for. A supporter buys AI capacity for a maintainer, and the maintainer spends it through whichever of the tools above they like. If that is you, the docs explain how a boost works, and you can create your boost page in a few minutes. If you want the longer argument for why maintainers should not be paying for this alone, I wrote it up earlier.
About the numbers
Stars, license, and owner type come from the GitHub REST API on October 22, 2026, and each group is ordered by stars. The descriptions come from each project's README, and the logos are the owners' GitHub avatars. Numbers move fast at this scale, so treat them as a snapshot. I plan to refresh them now and then and will update the date at the top when I do.