Best open-source AI coding agents (2026)
Ten agents you can read the source of, run yourself, and point at whichever model you like. No agent here is best at everything, so this is organised by what you are actually trying to do — and every infrastructure number comes from running these things on a real server rather than from a README.
Read this before the table
We sell hosting, and we offer one-click installs for five of the ten agents below — marked . That is a conflict of interest, so: the table is ordered by GitHub stars, never by whether we host something, and the agent we would point most people at first is one we do not host at all. Everything is sourced or measured, and where a project does not state something we leave the cell blank rather than filling it in.
The field
Ordered by GitHub stars. Popularity is not quality — it is a proxy for how easily you will find an answer when something breaks, which is worth something on its own.
| Agent | Stars | Licence | Language | Any model | MCP |
|---|---|---|---|---|---|
| OpenClaw | 379,000 | MIT | TypeScript | Yes | Native |
| OpenCode | 201,935 | MIT | TypeScript | — | — |
| Hermes Agent | 195,000 | MIT | Python | Yes | Native |
| Pi | 98,265 | MIT | TypeScript | Yes | — |
| Codex | 91,500 | Apache-2.0 | Rust | No | — |
| OpenHands | 77,400 | MIT | Python | Yes | Native |
| Cline | 63,400 | Apache-2.0 | TypeScript | Yes | Native |
| goose | 53,577 | Apache-2.0 | Rust | Yes | Native |
| Aider | 46,300 | Apache-2.0 | Python | Yes | — |
| Grok Build | 26,140 | Apache-2.0 | Rust | — | Native |
Licence, language and star counts read from each project’s repository. Blank cells mean the project does not state it — see the full matrix for every attribute and its source.
Picks, by what you are doing
Most people, most of the time → OpenCode
The largest open-source terminal agent by a wide margin, and it installs from practically any package manager you already have. It asks permission before running shell commands, which is the right default. We do not host it, and it is still where we would send someone with no other constraints.
You want governance you can trust → goose
Apache-2.0, built at Block, and now stewarded by the Linux Foundation. After Google retired Gemini CLI and replaced it with a closed-source successor, “who can switch this off?” stopped being a theoretical question — and a foundation is a materially different answer from a single vendor. It speaks MCP natively and runs against 15+ providers including local Ollama.
You want a web workspace, not a terminal → OpenHands
A browser-based workspace with an editor and a terminal, rather than a CLI. That makes it the most approachable agent here if you do not live in a shell — and the heaviest: we measured it at 1,151 MB on a 2 GB server, the only agent in this list that rules out the cheapest tier.
You want it running while you sleep → OpenClaw or Hermes
Both hold a gateway process open, so they can accept work on a schedule or over a channel rather than only when you type. That is a genuinely different shape from a CLI you invoke. Measured on a 1 vCPU / 1 GB server, OpenClaw settles at 278 MB and Hermes at 120 MB — Hermes is the lighter of the two by some distance, though its install takes about five minutes against OpenClaw’s one.
You live in an editor → Cline
Runs as a sidebar in VS Code, JetBrains and others, and now ships a CLI as well. If you want the agent where your code already is rather than in another window, this is the one.
You want small and sharp → Pi or Aider
Pi is a toolkit rather than a product — a unified provider API, an agent loop, a TUI library and a coding CLI, composable if you are building something of your own. Aider is the veteran git-native option: it commits as it works, which makes its changes trivial to review and undo.
What they cost to run
This is the part most roundups guess at. We provisioned real servers, installed each agent through our own one-click preset, and read memory thirty seconds after a live model request:
| Agent | Memory at idle | Proven on | Install |
|---|---|---|---|
| OpenClaw | 278 MB | 1 vCPU / 1 GB | under a minute |
| Hermes Agent | 120 MB | 1 vCPU / 1 GB | 5 minutes |
| Codex | nothing resident | 1 vCPU / 1 GB | under a minute |
| OpenHands | 846 MB | 1 vCPU / 2 GB | 3 minutes |
Two things fall out of this that are easy to miss. First, “CLI agent” hides two different shapes: Codex leaves nothing running between sessions, while OpenClaw and Hermes hold a process open — which is exactly why they can work while you are away and Codex cannot. Second, these are all far smaller than people assume, because the model runs on someone else’s hardware and your server is mostly waiting on the network.
The other six agents have no measured figures here, and we would rather say so than estimate them.
Method, and what this page does not tell you
- Open source only. Claude Code and Cursor CLI are excluded because they are closed, not because they are bad.
- Frameworks excluded. LangChain and similar are for building agents, not for using one.
- Sourced or measured. Public facts come from each repository; infrastructure figures were measured by us. Nothing is inferred to fill a gap.
- No quality benchmark yet. This page does not tell you which agent writes better code. Nobody has a trustworthy answer to that, ourselves included — and a roundup that pretends otherwise is guessing at the most important question.
- Star counts move. Each figure carries the date it was read; check the repository if the number matters to your decision.
Related
- VPS hosting for AI agents — the full attribute matrix, with sources per cell.
- Gemini CLI was retired — why governance is now a selection criterion.
- Managed vs self-hosted — including the evidence against self-hosting.