gem-hunter
主要言語: Python
Curated directory of terminal-native AI coding agents and the harnesses that orchestrate them. Covers open-source tools (Pi, OpenCode, Aider, Goose), platform agents (Claude Code, Codex, Gemini CLI), parallel runners, autonomous loops, and agent infrastructure.
A curated list of 150+ CLI coding agents — AI-powered tools that live in your terminal, read/edit repos, and run commands — plus the harnesses that orchestrate, sandbox, or extend them.
Last updated: 2026-10-26
A CLI coding agent is an AI-powered tool that runs in your terminal and can autonomously read, write, and execute code in your repository. Unlike chat-based assistants, these agents have direct access to your filesystem, shell, and dev tools — they can edit files, run tests, commit changes, and iterate on errors. Think of them as AI pair programmers that live where you already work: the command line.
Forkable, extensible, and community-driven. Sorted by GitHub stars. Provider tags [Company] indicate the backing organization.
Hermes Agent ⭐ 257k — Nous Research's self-improving CLI agent with persistent memory, automated skill creation, sandboxed code execution via Unix socket RPC, and multi-platform reach (Telegram/Slack/Discord/WhatsApp); supports 300+ models across multiple providers.
OpenCode ⭐ 222k — Terminal-native coding agent with 75+ provider support, LSP integration, and privacy-first design (formerly opencode-ai; now at opencode.ai).
Claw Code ⭐ 195k — Clean-room Python/Rust rewrite of Claude Code architecture using oh-my-codex; fastest repo in GitHub history to 100K stars. Born from the March 2026 Claude Code source leak. MIT.
Codex CLI ⭐ 131k [OpenAI] — OpenAI's local coding agent for reading/editing/running code, with an interactive TUI and tool execution. Apache-2.0.
Pi ⭐ 118k — Minimal, adaptable terminal coding harness from the pi-mono toolkit; unified LLM API, TUI, skills, and MCP support.
Gemini CLI ⭐ 108k [Google] — Google's terminal agent powered by Gemini, with tools for repo work and research. Apache-2.0.
OpenHands ⭐ 91.9k — Open-source agentic developer environment (formerly OpenDevin) with CLI and web entrypoints; also has a lightweight CLI-only package.
Cline CLI ⭐ 70.9k — Model-agnostic autonomous agent for planning, file edits, command execution, and browser use.
Oh My OpenAgent ⭐ 70.3k — Multi-harness agent OS layered over OpenCode, Codex, and Pi; Team Mode runs several models on one job, with a hash-anchored tool harness and heavy context tuning. npm oh-my-opencode; the Codex-only variant installs via npx lazycodex-ai install. Source-available (SUL-1.0, not OSI).
Open Interpreter ⭐ 68.7k — Terminal tool that can execute code and actions; often used as a "do things on my machine" agent.
Goose ⭐ 55.8k — Local, extensible agent that can execute, edit, and test; designed to run on-device and integrate with MCP.
Aider ⭐ 49.8k — Pair-programming agent for editing files via diffs/patches, with strong git and multi-file workflows.
Codewhale ⭐ 41.1k — Rust terminal coding agent (formerly deepseek-tui), bring-your-own-model across 30+ providers plus local vLLM/SGLang/Ollama. TUI or headless codewhale exec for scripts and CI; fleets pin a different provider, model, and reasoning tier per role, so a cheap model can direct an expensive one. Workspace snapshots, /undo, and a three-level permission posture. MIT.
OH-MY-PI ⭐ 36.7k — Terminal coding agent ("Pi") with a TypeScript/Rust monorepo and local-first ergonomics.
Continue CLI ⭐ 36.3k — Open-source terminal extension for multi-model coding with local/privacy focus.
Reasonix ⭐ 36k — Single Go binary coding agent built for long unattended runs; one local engine with four ways in (terminal, desktop app, browser, or editor over ACP). Config-driven providers in reasonix.toml, optional split executor/planner models, MCP plus an Extension Protocol sidecar system, plan mode, workspace sandbox, and per-turn checkpoints. MIT.
Deep Agents Code ⭐ 30.6k [LangChain] — LangChain's official terminal coding agent built on the Deep Agents SDK; interactive TUI, file ops, shell access, subagents, headless mode, and human-in-the-loop approvals with any tool-calling LLM. PyPI deepagents-code.
Crush ⭐ 28.8k — Charmbracelet's glamorous agentic coding TUI in Go; multi-provider, LSP-aware, with rich terminal UI.
Qwen Code ⭐ 28.7k [Alibaba] — Alibaba Qwen's official CLI agent for Qwen coder models (workflow tool + repo operations). Apache-2.0.
Kilo Code CLI ⭐ 27.8k — Agentic engineering platform with CLI; orchestrator mode, 100s of LLMs, skills, and checkpointing.
Grok Build ⭐ 27.6k [xAI] — xAI's official coding agent harness and TUI; fullscreen, mouse-interactive, and extensible. Apache-2.0.
Roo Code CLI ⭐ 24.3k — Multi-mode CLI agent (architect/code/debug/orchestrator modes); Claude-like terminal interface with skills and checkpoints.
Prime Agent ⭐ 22.3k [Prime Intellect] — Self-improving RLM coding agent where a persistent IPython kernel is the model's only tool, so file edits, shell commands, skills, and subagents (rlm(...)) all happen as Python. A "Continual Harness" keeps memories, skills, and subagent specs as durable state that /refine updates from session evidence; daemon-backed sessions survive terminal disconnect, with goals, heartbeats, schedules, and bounded autonomous mode. Built on Pi. MIT.
jcode ⭐ 21.2k — Rust TUI agent optimized for RAM and startup latency (~28 MB PSS per session with local embeddings off), built for scaling many parallel sessions. Agent memory, swarm mode, browser automation, MCP, and 40+ providers with built-in OAuth login flows. MIT.
SWE-agent ⭐ 20.6k — Agent for resolving real repo issues/PR tasks; frequently used in SWE-bench-style workflows.
Plandex ⭐ 15.8k — "Plan-first" CLI agent for building features across multiple files with structured steps and 2M token context.
Codebuff ⭐ 13.8k — Multi-agent AI coding assistant with CLI support for collaborative coding workflows.
MiMo Code ⭐ 13.7k [Xiaomi] — Xiaomi's official terminal coding agent; TUI + non-interactive modes, MCP, skills, hooks, git worktrees, and resumable sessions. Runs MiMo-V2.5-Pro or any mainstream provider; npm @xiaomi-mimo/cli. MIT.
Smol Developer ⭐ 12.2k — Embeddable developer agent that generates entire codebases from a prompt; designed to be embedded in apps.
Trae Agent ⭐ 12.2k [ByteDance] — ByteDance's research-friendly CLI agent for software engineering tasks, with modular architecture and multi-LLM support. MIT.
Kimi CLI ⭐ 11.4k [Moonshot AI] — Moonshot AI's CLI coding agent with skills, MCP support, and ACP IDE integration; being wound down in favor of Kimi Code. Apache-2.0.
Claude Engineer ⭐ 11.2k — Community-driven CLI for agentic Claude workflows with file management and iterative development.
Claurst ⭐ 10.3k — Claude Code rewritten in idiomatic Rust with architectural breakdown; includes discoveries from the source leak (KAIROS persistent assistant, buddy system). MIT.
Free Code ⭐ 8.8k — Fork of Claude Code with all telemetry removed, guardrails stripped, and all experimental features enabled (KAIROS, dream mode, companion system).
Kimi Code ⭐ 8k [Moonshot AI] — Moonshot AI's next-gen TypeScript coding agent — a separate project from Kimi CLI. Event-sourced agent runtime with replayable state, context undo/compaction, and blob offloading; goal mode for bounded autonomous runs, background subagents, skills plus plugin manifests (prompts/MCP servers/commands), and subscription OAuth. npm @moonshot-ai/kimi-code. MIT.
ForgeCode ⭐ 7.6k — AI pair programmer supporting 300+ models, with task management, custom agents, and large-scale refactor tooling.
OpenSquilla ⭐ 7.1k — Self-hostable microkernel agent runtime with a full CLI (opensquilla chat REPL, one-shot agent mode, gateway); autonomous file edits, shell and background processes, git tools, ML-based tier routing, sandboxing (Bubblewrap/Seatbelt), persistent memory, and 20+ providers. Apache-2.0.
ANUS ⭐ 6.6k — Free coding agent built on Pi: every request goes to the smartest free model that is answering (OpenRouter, Gemini, Groq, Mistral free tiers), and the next one takes over when a model hits its limit. The repository's star count predates the agent — it was reused from an unrelated 2025 project. MIT.
Kode CLI ⭐ 5.2k — ShareAI's open-source CLI agent for terminal-native coding with multi-provider support.
Mistral Vibe ⭐ 5.2k [Mistral] — Mistral's CLI coding assistant for conversational repo interaction and edits. Apache-2.0.
gptme ⭐ 4.5k — AI agent in your terminal with support for persistent autonomous agents. Runs code, edits files, browses the web. Build long-lived self-modifying agents with git-backed memory via gptme-agent-template.
Every Code ⭐ 4k — Fork of Codex CLI with validation, automation, browser integration, multi-agents, theming, and multi-provider orchestration (OpenAI, Claude, Gemini). Apache-2.0.
Letta Code ⭐ 3.7k — Memory-first CLI coding agent built on the Letta platform (formerly MemGPT); persistent memory across sessions, model-agnostic (Claude/GPT/Gemini), skill learning, and context repositories.
Grok CLI ⭐ 3.5k — Community CLI agent built on xAI's Grok models for terminal-based coding tasks.
Devon ⭐ 3.5k — Open-source pair programmer with a TUI; autonomous planning, execution, and debugging in Git workflows.
Atomic Agent ⭐ 3.4k — Local-first agent that runs the control loop on your machine: edits files, runs approved shell commands, drives your browser, and carries context across sessions, on local llama.cpp or cloud models. Interactive TUI, non-interactive run, and an OpenAI-compatible serve mode. Node.js, MIT.
BitFun ⭐ 3.3k — Rust coding agent with an interactive TUI and non-interactive bitfun exec runs; session fork/undo/redo/compact with a timeline, MCP and ACP, skills, hooks, and daemon-backed background tasks. Ships alongside, but builds independently of, a Tauri desktop app. MIT.
AutoCodeRover ⭐ 3.1k — Autonomous program improvement agent; patches real GitHub issues using code search and analysis.
Tau ⭐ 3k [Hugging Face] — Small, readable Python coding agent in the terminal, inspired by Pi; CLI and Textual TUI over a provider-neutral core (Anthropic, Google, Mistral, OpenAI-compatible, Codex), with file/shell tools, durable JSONL sessions with branching, skills, and MCP-style extensions. Doubles as a teaching codebase for how agent harnesses are built. PyPI tau-ai. MIT.
Nanocoder ⭐ 2.5k — Local-first CLI coding agent built by a community collective; bring your own model (Ollama, OpenRouter, or any OpenAI-compatible API), native tool calling with XML fallback, MCP support, and file-based custom commands and tools. MIT.
CodeMachine-CLI ⭐ 2.5k — Community multi-agent CLI aimed at running coding workflows locally (vibe-coding oriented).
Codel ⭐ 2.5k — Autonomous agent for performing complex tasks via terminal; runs in Docker with a web UI.
open-codex ⭐ 2.4k — Lightweight fork of Codex CLI with multi-provider support (OpenAI, Gemini, OpenRouter, Ollama).
RA.Aid ⭐ 2.2k — Autonomous coding agent built on LangGraph with research/plan/implement pipeline; optional aider integration for near-full autonomy.
Agentless ⭐ 2.1k — Lightweight approach to autonomous software engineering without persistent agent loops.
Amazon Q Developer CLI ⭐ 2k [AWS] — AWS's agentic terminal chat for building apps, debugging, and DevOps with natural language. Apache-2.0.
Ouroboros ⭐ 1.6k — Self-evolving general agent with durable memory and identity across restarts; coordinates a subagent swarm and can rewrite its own implementation behind a review gate. Native desktop app or headless ouroboros run/tasks/chat/logs/evolve against a local gateway. Python, MIT.
Neovate Code ⭐ 1.6k [Ant Group] — Ant Group's CLI agent with plugin system, multi-model/multi-provider support, MCP integrations, and headless automation mode. MIT.
Empryo ⭐ 1.5k — Coding harness (formerly SoulForge) that maps the repo before it edits (tree-sitter symbols, import graph, git co-change) and edits by symbol through AST and LSP instead of find and replace, then runs typecheck, lint and tests. Terminal UI, desktop app and headless CLI share one engine and session; many providers or ChatGPT/Copilot/Gemini plan sign-in, per-task model routing, parallel subagents, MCP and skills. Source-available (BUSL-1.1).
Maki ⭐ 1.3k — Rust TUI coding agent optimized for context-token efficiency: tree-sitter file indexing, a sandboxed code_execution tool that chains tool calls without polluting the context window, tree-sitter-parsed permission gating for bash commands, and a memory tool. Extensible via neovim-like Lua plugins; 17+ providers, skills, MCP, ACP, and a Claude Code-compatible headless mode. MIT.
Tianshu Harness ⭐ 1.2k — TypeScript coding-agent runtime with one kernel for a terminal TUI and a desktop GUI. The harness externalizes goals, evidence, and completion so a model claim of done is not treated as done, with multi-agent orchestration and prefix-cache-friendly long sessions. Apache-2.0.
Pydantic Deep Agents ⭐ 1.1k — Terminal coding agent built on Pydantic AI that edits files, runs shell commands, plans, delegates to sub-agents and connects to MCP servers, with optional Docker-sandboxed execution; works with any model, including local ones through Ollama. The same harness is available as a Python library (pydantic-deep). MIT.
hax ⭐ 937 — Minimalist terminal-native coding agent written in C; a single native binary that starts instantly and uses a few MB of RAM, leaving the machine's memory to a local model. First-class llama.cpp support (auto-discovers the served model and its runtime capabilities, no provider config) alongside OpenAI- and Anthropic-compatible endpoints, Codex via a ChatGPT subscription, and OpenRouter; streaming Markdown TUI that preserves native scrollback and a clean one-shot -p mode. Linux/macOS/BSD. MIT.
VT Code ⭐ 890 — Open-source coding agent with LLM-native code understanding and robust shell safety. Supports multiple LLM providers with automatic failover and efficient context management. MIT.
Groq Code CLI ⭐ 745 — Customizable, lightweight CLI powered by Groq's ultra-fast inference; extensible tools/commands with multi-model support.
Tura ⭐ 657 — Rust agent runtime that compiles a task into a single runtime-managed command graph instead of a model round trip per tool call, with backward planning and task-scoped compaction. Vendor-published DeepSWE comparison claims 77.5% fewer tokens than Codex CLI. TUI plus a Tauri GUI; npm tura-ai. AGPL-3.0.
Dexto ⭐ 657 — Coding agent and general agent harness with CLI/web/API modes; ships a production-ready coding agent with sub-agent spawning.
agentty ⭐ 617 — Native C++26 terminal coding agent and drop-in claude-code alternative; single ~13.6 MB static binary with sub-millisecond cold start and zero runtime deps (no Node/Python/Electron). Sandboxed by default (Bubblewrap/sandbox-exec), model-agnostic (Claude, OpenAI, Groq, OpenRouter, Together, Cerebras, local Ollama), runs inside Zed over ACP, and drives air-gapped hosts over SSH. Linux/macOS/Windows/Termux. MIT.
mu ⭐ 552 — Coding agent built on Pi with a judgment kernel: a small, fast judge (Jev, a local model, or any LLM) makes the routine calls at 38 decision points (what enters the context, whether a flagged command was asked for, prompt-injection screening, whether "done" was verified), so the main model keeps its attention for the work. Sub-agents, a built-in browser, and a desktop app. npm mu-agent. MIT; the desktop app is Apache-2.0.
claw-code-agent ⭐ 546 — Python-only Claude Code rewrite with zero external dependencies; born from the March 2026 Claude Code source leak, positioned as easier to hack on than the Rust/TypeScript reimplementations.
g3 ⭐ 520 — "Coding AI agent" in Rust: tool-running, repo interaction, skills system, and provider abstraction.
Orca ⭐ 510 — DeepSeek-native terminal coding agent in Rust with OS-level sandboxing (Seatbelt, bwrap, Landlock+seccomp, fail-closed), persistent goal mode with stall detection, background tasks, JS workflows, folder trust, and 1M-context auto-compaction. Single binary. MIT.
cecli ⭐ 415 — Community fork of Aider (formerly aider-ce) that has stayed in active development since upstream slowed, adding extensibility and customization hooks to the pair-programming CLI. Python, Apache-2.0.
Coro Code ⭐ 369 — Open-source CLI coding agent, a free alternative to Claude Code; generate, debug, and manage code seamlessly.
zot ⭐ 351 — Zero-overhead and lightweight coding agent harness with TUI/JSON/RPC modes, structured tools, reviewable file diffs, skills, extensions, and optional guardrails.
LettaBot ⭐ 326 — Personal AI assistant with persistent unified memory across Telegram, Slack, Discord, WhatsApp, and Signal; built on the Letta platform.
Mini-Kode ⭐ 307 — An educational AI coding agent CLI, intended as a readable reference implementation.
Mixdog ⭐ 199 — Desktop app and terminal coding agent (npm install -g mixdog) that runs on an existing Claude or ChatGPT subscription and keeps per-turn context small: tools and skills load on demand and long outputs are offloaded to files. The project's own Terminal-Bench 2.1 run reports 63% fewer tokens than Codex CLI; logs in repo. Apache-2.0.
nori-cli ⭐ 191 — Multi-provider CLI built on Codex CLI; switch between Claude, Gemini, and Codex from the same native terminal.
VibePod ⭐ 187 — Unified CLI for running AI coding agents in isolated Docker containers; zero-config setup, local metrics, HTTP traffic tracking, and an analytics dashboard for side-by-side comparison.
Octomind ⭐ 179 — Token-efficient, model-agnostic coding agent runtime in one Rust binary: interactive, piped, daemon, WebSocket, or ACP sub-agent. On its own octobench suite it reports 24/25 real-PR tasks solved at lower cost than Claude Code and opencode; adaptive cache-aware compaction, guardrails-as-code, MCP-native, shareable specialist taps. Apache-2.0.
rpi ⭐ 154 — Library-first Rust coding agent, inspired by Pi but independent of it: nine composable crates (rpi-agent, rpi-tools, rpi-harness, rpi-tui, …) so the agent loop can be embedded in your own program, plus a ready-to-run rpi terminal CLI. Crash-resumable JSONL sessions that self-heal a duplicated-seq log, a stable #[repr(C)] plugin ABI for tools and providers, and a provider-agnostic StreamFn boundary that makes the loop testable with no network. MIT.
San ⭐ 144 — Go terminal-native runtime for specialized AI agents; provider-agnostic (Anthropic, OpenAI, Google, DeepSeek, Moonshot, Qwen, MiniMax, GLM), runs Claude Code skills/plugins/MCP unmodified, swappable search backends, custom personas, sandboxed subagents, lifecycle hooks, and a self-evolving memory loop. Single ~12 MB binary, zero runtime deps. Apache-2.0.
San ⭐ 144 — Go terminal-native runtime for specialized AI agents; provider-agnostic (Anthropic, OpenAI, Google, DeepSeek, Moonshot, Qwen, MiniMax, GLM), runs Claude Code skills/plugins/MCP unmodified, swappable search backends, custom personas, sandboxed subagents, lifecycle hooks, and a self-evolving memory loop. Single ~12 MB binary, zero runtime deps. Apache-2.0.
cursor-agent ⭐ 137 — Python-based agent replicating Cursor's coding assistant capabilities; supports Claude, OpenAI, and local Ollama models.
Waveloom ⭐ 136 — Go terminal-native coding agent with Bubble Tea TUI; DeepSeek V4 prompt caching for long-context efficiency; Claude Code-compatible UX with skill/MCP auto-discovery; four-tier context compaction, three subagent modes (Fork/Cold/Explore), permission engine, and plan mode. Single ~19 MB binary, zero runtime deps. Apache-2.0.
DvalinCode ⭐ 118 — Provider-neutral, local-first coding agent (Chat/Cowork/Code modes) built for governance: an org policy engine, enforced network egress (per-request checks plus OS-sandboxed subprocesses via sandbox-exec/Bubblewrap), and a tamper-evident, hash-chained audit trail. Inline diff approval, durable session journal, built-in Web GUI from a single binary; works with any OpenAI-compatible endpoint (DeepSeek, OpenAI, Claude via OpenRouter, Groq, Ollama). Zero runtime deps. MIT.
3code ⭐ 113 — Free and open source command-line coding agent built from the ground up to be efficient enough to use 3rd party token providers without a coding plan. Tight, no-frills interface, instant startup, works on macOS, Windows, Linux, and Termux; supports a wide range of providers including EU ones (Mistral, TensorX). MIT.
Codex Infinity ⭐ 102 — Autonomous terminal coding agent (OpenAI Codex CLI fork) adding auto-continuation, parallel multi-agent runs, and CI repair loops.
openHarness ⭐ 100 — Open-source Claude Code alternative. 78 slash commands, 42 tools, MCP (stdio/HTTP/SSE + OAuth 2.1), hooks, subagents, plan mode. Works with Anthropic/OpenAI/Ollama/llama.cpp/LM Studio. Ships both npm and Python SDK. MIT.
Martty ⭐ 80 — Rust/ratatui terminal ACP client that bundles its own ACP layer, so agents that speak the protocol run in a native TUI rather than each shipping their own. Session history, MCP, and a martty single binary. MIT.
Crab Code ⭐ 77 — Rust-native coding agent built from scratch, deliberately mirroring Claude Code's toolset, permission model, and interaction patterns; Anthropic, OpenAI, DeepSeek, Bedrock, and Vertex providers. MIT.
AgenC Core ⭐ 75 — Open source terminal coding agent that works with any model provider, including local models through Ollama and LM Studio. TUI, headless mode, background agents, MCP. Self-reported 76.2% on Terminal-Bench 2.1 for $34 with DeepSeek V4.1 Flash. MIT.
Keen Code ⭐ 74 — Go-based CLI coding agent focused on efficient context management — uses lean TurnMemory summaries instead of raw tool traces to maximize context efficiency. 9+ providers, skill-driven MCP servers, and full transparency with every prompt, design doc, and implementation plan saved as markdown. MIT.
picocode ⭐ 63 — Minimal Rust-based coding agent focused on CI workflows and small codemods; multi-LLM with personas.
Smelt ⭐ 57 — Rust TUI coding agent; multi-provider (Anthropic, OpenAI, Ollama, GitHub Copilot, any OpenAI-compatible endpoint), four modes (Normal/Plan/Apply/Yolo), granular permission system, parallel subagents, vim keybindings, and headless scriptable mode. MIT.
Jazz ⭐ 54 — Generalist agent harness that runs your agent anywhere: terminal, scripts, cron, a GitHub Action that reviews PRs, or a Telegram/Discord bot you own. Define model, persona, toolset, and permissions in one JSON file; 18 providers (incl. local Ollama/llama.cpp) plus MCP, with permission-gated file, git, and web tools that ask for approval wherever you are. MIT.
QQCode ⭐ 54 — Lightweight CLI coding agent in Rust focused on speed, determinism, and developer control; supports skills.
opencode-rust ⭐ 52 — Independent Rust reimplementation of OpenCode, not affiliated with the OpenCode team: CLI, TUI and HTTP server for local AI agents across 21 crates, no Docker required. npm opencode-rust, or brew install vakandi/tap/opencode-rust for the opr binary. MIT.
Zap ⭐ 36 — Skill-first Rust TUI coding agent that injects only the context your task needs — no system prompt bloat. Single binary, no runtime. Supports Claude, Gemini, OpenAI, and local models via LM Studio; code-indexed via SQLite for fast symbol lookup; MCP support. MIT.
Nausicaa ⭐ 33 — TypeScript CLI/runtime for general-purpose agent tasks with addressable Lanes, an optional Teto observer lane, durable Run ledgers, daemon/worker lifecycle, and cross-Run A2A. npm nausicaa-harness, MIT.
Grinta ⭐ 32 — Local-first, provider-agnostic terminal coding agent built for long-horizon autonomous execution; durable state and recovery, context management, structured tool orchestration, LSP/DAP integration, and validation-gated completion. Python, MIT.
Hazzel ⭐ 28 — Approval-first Python coding agent where lookup-shaped asks ("find all *.py files") are answered locally with no model round trip, and /usage//budget report provider-billed dollars rather than estimates. Every write is a diff you approve, and /undo reverts only the agent's hunks while keeping edits you made afterwards. ~10k lines readable end to end, 8 bring-your-own-key providers including Ollama, MCP client, Apache-2.0.
OB-1 ⭐ 26 — Terminal coding agent that runs with no account, API key, or card against a free-model catalog. Persistent project memory builds a fact and relationship graph from real work and surfaces it with /memory. npm ob1. Apache-2.0.
Kolega Code ⭐ 21 — Python terminal coding agent where the model writes its own multi-agent workflows (Gigacode): the runtime enforces concurrency caps and token budgets, journals every completed agent call to disk, and replays unchanged calls for free on resume. 15+ providers, MCP over stdio/SSE/streamable HTTP, ACP, sandboxed shell and filesystem tools, git worktrees, and a browser agent. Source-available (BUSL-1.1).
Mercury CLI ⭐ 21 — Multi-session terminal coding agent (mercury-cli.ai): runs several coding sessions side by side on one screen, spawns crewmates on any of 13 providers or a local model in isolated git worktrees with file leases, and keeps project memory across sessions (BM25 plus an on-device embedding model). LSP, 23-language AST rewrites, a DAP debugger, keyless web search, computer use and voice dictation built in. npm mercury-tech-cli, TypeScript. Source-available (BSL 1.1).
Binharic ⭐ 19 — A multi-provider "tech-priest persona" coding agent CLI (stylized, tool-using).
memcode ⭐ 18 — Go terminal coding agent that keeps persistent per-project memory in a local .memcode store — subsystems, prior work, failed approaches, corrected preferences — so sessions don't start from zero. Single binary that doubles as a self-hosted gateway answering on twelve chat channels, and ships an MCP memory server so other agents can read the same store. Runs on your own API keys, a local endpoint like Ollama, or a hosted account. MIT.
Darce ⭐ 11 — Ultralight (14 kB) multi-model CLI agent built with Ink; 7 tools, smart model routing across providers, streaming, session resume, and slash commands. MIT.
slash-agent ⭐ 10 — Shell-integrated terminal copilot for active Bash/Zsh/Fish shells; captures tmux/history scrollback to diagnose errors, runs commands via interactive PTY bridge with shell state sync (cd/export), and supports local Ollama/OpenAI via py-agent-core. Apache-2.0.
Forge (Norvia Labs) ⭐ 8 — Rust TUI coding agent unifying an AI agent, a vim-style code editor, and a shell in one keyboard-driven workspace; approval-aware command execution, durable SQLite session journal, and MCP server support. MIT.
CLAII ⭐ 6 — CLI-first AI coding agent with multi-agent orchestration, MCP toolchains, and memory-persistent refactors.
Forge Harness ⭐ 5 — Local-first TypeScript coding-agent CLI optimized for local and small language models; interactive chat and headless runs with bounded repo navigation, approval-gated transactional edits, command execution + verification, durable sessions/undo, and isolated worktree execution. npm @aglegg/forge-harness. Apache-2.0.
AshlrCode ⭐ 5 — Multi-provider terminal coding agent (ac) on Bun, defaulting to xAI Grok with failover to Anthropic, OpenAI, DeepSeek, Groq and Ollama; built-in tools, skills, sub-agents, permission prompts with regex allow/deny rules, MCP over stdio/SSE/WebSocket, and an autopilot mode. npm ashlrcode. MIT.
ipsupport-code ⭐ 4 — Go terminal coding agent built for LM Studio and other OpenAI-compatible endpoints, single static binary; a small fat-tool schema tuned for weak local models, opt-in OS-level sandboxing (Seatbelt on macOS, Landlock on Linux) for its shell tool, and a reflect step after each task that writes new lessons to disk. MIT.
Ferrum ⭐ 4 — Small Linux-only Rust-native coding agent with interactive and headless modes, ACP, safety-tiered native tools, durable JSONL sessions, Codex/ChatGPT OAuth, OpenAI-compatible providers, MCP, skills, and image input. MIT; primary development is on Codeberg.
Memdoor ⭐ 4 — Go terminal coding agent on your own API key or a ChatGPT plan. Describe the steps and it writes them as a workflow that runs as a graph: a step is done only when its file exists or its command exits 0, with approval gates and resume; a decision model in front of the chat model cuts input tokens by the author's own measurements. Apache-2.0.
MoCode ⭐ 3 — Terminal coding agent (TypeScript) with durable lifecycle memory, hash-gated skill trust, and a built-in 61-task eval harness; drives any OpenAI-compatible backend (GLM, DeepSeek, Qwen, Ollama, vLLM) without a vendor key. npm mocode-ai. MIT.
WorkGround2 ⭐ 2 — Local-first AI engineering workbench: one Go agent kernel behind CLI/TUI, web serve, a Wails desktop app, and IM bots (Feishu/WeCom/QQ), with multi-model providers, MCP + plugins, project memory, sandboxed/approved execution, and checkpoints + /rewind. MIT.
Kolkrabbi ⭐ 2 — Go terminal coding agent (kolk) whose effort dial — low to ultra — picks which model runs each turn, not one model's thinking budget. Local models on this machine, the LAN, or one address you name; any OpenAI-compatible endpoint; or your Claude and ChatGPT subscriptions through each vendor's own CLI, no credential held. Checkpoints every write and logs each call's cost. Single static binary, Apache-2.0.
ucode ⭐ 2 — Free terminal coding agent on Google Gemini's free tier (or Ollama, offline): scaffolds from a working starter, then opens and clicks the page to check its own work; type checks and related tests after every edit, shadow-git /undo, MCP, hooks, skills, a headless -p --json mode and voice input. npm ucode-agent. AGPL-3.0.
one-agent ⭐ 2 — Autonomous Python terminal agent (CLI + TUI + JSON-RPC); supports headless/RE