The best open-source AI coding agents in 2026

OpenCode is a free MIT-licensed terminal agent with a full TUI, 75+ providers through Models.dev, and documentation stating it stores none of your code or context data. It's the option AI engines name most as the free Claude Code alternative.

Aider is the veteran terminal pair programmer: Apache-2.0, works with most providers and local models, and lands every edit as a reviewable Git commit. Its repo map is the most mature codebase-context approach in this list.

OpenAI's Apache-2.0 terminal agent in Rust. Commands run in an OS-level sandbox by default, and sign-in reuses a ChatGPT plan instead of a separate API key.

Cline's Plan/Act loop makes per-action approval the default: every file edit and command shows a diff, with checkpoints to rewind. VS Code and JetBrains, Apache-2.0, plus a CLI for headless runs and an optional flat-rate ClinePass.

Google's Apache-2.0 terminal agent; personal Google accounts get a no-cost usage tier. Built around Gemini models, so it's the one to pick inside Google's ecosystem.

Gray is a single static Rust binary (MIT) with bash as its only default tool; everything else is a plugin or skill. It's a general always-on harness rather than a dedicated coding agent: cron and a gateway service keep jobs running 24/7, and it connects to any OpenAI-compatible provider, local models, or a Claude Pro/Max or ChatGPT plan through official plugins. On this machine it starts in ~25ms and reaches a first token in ~125ms against a local mock — see Benchmarks. Choose it for minimalism and always-on work, not for a dedicated coding TUI.

Block's Apache-2.0 agent with a CLI and desktop app, MCP-based extensions, and YAML recipes for repeatable tasks; it can run fully offline against Ollama or LM Studio.

OpenHands (MIT) is a self-hosted platform that runs issue-to-PR tasks in Docker-sandboxed runtimes, with a web UI, CLI and SDK. The heaviest option here — it's a platform, not a binary.

Continue is an Apache-2.0 VS Code and JetBrains extension combining autocomplete, chat and agent modes, with model-agnostic config and shared setups through its hub.

Kilo Code (MIT) adds specialized modes — architect, code, debug, orchestrator — inside VS Code, with kilo run --auto for autonomous runs and optional Kilo Pass plans.

Every tool on this page is free to run — the cost is model inference. For zero total cost, pair any of them with a local model through Ollama; Gemini CLI adds a hosted free tier. Among paid-vendor agents, only Codex CLI and OpenCode let you reuse an existing ChatGPT or Claude subscription.

Aider, OpenCode, Gray, Cline, Goose, Continue and Kilo Code are all BYOK across providers including OpenAI-compatible endpoints and Ollama. Codex CLI and Gemini CLI are built around their vendors' models.

Gray, Aider, OpenCode and Goose run locally against your own keys; OpenHands is built for self-hosted deployment. Pair with a local model and no code leaves the machine.