Compare

Coding agents

The AI tools that actually write and ship code — from cloud coding agents to free, open-source options you run yourself, and where your code goes in each.

↻ Checked by Hermes · 15 June 2026
The Quick Version
  • Claude Code and Cursor lead — Claude Code for terminal and agentic multi-file work, Cursor for the IDE experience.
  • Strong free, open-source options (Cline, OpenHands, Aider) let you bring your own key and control the cost.
  • Where your code goes splits the field: cloud agents send it to a vendor; bring-your-own-key tools send it only to the model API you choose.
  • Open-source picks need a model API key; the paid ones bundle a subscription.
  • Scores are RunAgentRun's editorial view; the factual fields are re-checked daily.

AI that writes and ships real code has moved fast. The split now is between polished paid agents and capable open-source tools you run yourself with your own model key. Beyond capability, the deciding question for many teams is where your code goes — to a vendor, or only to a model API you control. Here's how they compare for a small team.

The ranking

01
Claude Code

Agentic coding in the terminal and IDE — strong at multi-file changes and following a plan.

Type: CLI · IDEOpen source: NoModels: ClaudePricing: Subscription / usage
4.5
★★★★
Visit ↗
02
Cursor

The AI-native IDE most teams start with — fast inline edits and a capable agent mode.

Type: IDEOpen source: NoModels: Multi-modelPricing: ~$20/mo
4.3
★★★★
Visit ↗
03
OpenAI Codex

OpenAI's agentic coder — give it a task and it works the repo; tightest in the OpenAI stack.

Type: CLI · cloudOpen source: NoModels: OpenAIPricing: Subscription / usage
4.1
★★★★
Visit ↗
04
Cline

Open-source VS Code agent — bring your own API key and keep full control of cost and models.

Type: IDE (VS Code)Open source: Yes (Apache-2.0)Models: Any (BYO key)Pricing: Free + model cost
4.0
★★★★
Visit ↗
05
OpenHands

Open-source autonomous developer you can self-host — the most hands-off of the open options.

Type: Autonomous · self-hostOpen source: Yes (MIT)Models: Any (BYO key)Pricing: Free + model cost
3.8
★★★★
Visit ↗
06
Aider

Open-source CLI pair-programmer — lightweight, git-native, brilliant for focused changes.

Type: CLIOpen source: Yes (Apache-2.0)Models: Any (BYO key)Pricing: Free + model cost
3.7
★★★
Visit ↗
07
Kimi (Moonshot)

Long-context coding and agent work at aggressive pricing — the value pick from the open-weight camp.

Type: Web · agentOpen source: Weights openModels: KimiPricing: Usage
3.5
★★★
Visit ↗

Side by side

scroll across →

Every consideration at a glance. Scores are our editorial view; the rest are facts, re-checked daily.

Tool Score TypeOpen sourceModelsYour codePricing
Claude Code ↗ 4.5 CLI · IDENoClaudeTo Anthropic API (not trained on)Subscription / usage
Cursor ↗ 4.3 IDENoMulti-modelTo Cursor + model (Privacy Mode)~$20/mo
OpenAI Codex ↗ 4.1 CLI · cloudNoOpenAITo OpenAI (not trained on)Subscription / usage
Cline ↗ 4.0 IDE (VS Code)Yes (Apache-2.0)Any (BYO key)Only to your model APIFree + model cost
OpenHands ↗ 3.8 Autonomous · self-hostYes (MIT)Any (BYO key)Self-hosted / your APIFree + model cost
Aider ↗ 3.7 CLIYes (Apache-2.0)Any (BYO key)Only to your model APIFree + model cost
Kimi (Moonshot) ↗ 3.5 Web · agentWeights openKimiTo Moonshot (or self-host)Usage

In detail

Claude Code · 4.5/5

Agentic coding in the terminal and IDE — strong at planning and multi-file changes, with a human approving before anything ships. It runs on your machine and sends code to Anthropic's API, which doesn't train on commercial usage. Subscription or usage-based, and the tool many of these comparisons are now measured against.

Type: CLI · IDE · Open source: No · Models: Claude · Your code: To Anthropic API (not trained on) · Pricing: Subscription / usage · visit ↗

Cursor · 4.3/5

The AI-native IDE most teams reach for first — fast inline edits and a capable agent mode, at around $20/mo per seat. Privacy Mode keeps your code from being stored or trained on. The gentlest on-ramp if your team wants a familiar editor.

Type: IDE · Open source: No · Models: Multi-model · Your code: To Cursor + model (Privacy Mode) · Pricing: ~$20/mo · visit ↗

OpenAI Codex · 4.1/5

Give it a task and it works the repo on its own — tightest fit if you're already in the OpenAI ecosystem. Code is sent to OpenAI, which doesn't train on API/business usage by default. Subscription or usage-based.

Type: CLI · cloud · Open source: No · Models: OpenAI · Your code: To OpenAI (not trained on) · Pricing: Subscription / usage · visit ↗

Cline · 4.0/5

An open-source VS Code agent where you bring your own model API key — so your code only ever goes to the model provider you choose, and you keep full control of cost. The best open option if your team already lives in VS Code.

Type: IDE (VS Code) · Open source: Yes (Apache-2.0) · Models: Any (BYO key) · Your code: Only to your model API · Pricing: Free + model cost · visit ↗

OpenHands · 3.8/5

An open-source autonomous developer you can self-host — the most hands-off of the open options, MIT-licensed. Run it on your own infrastructure and your code never leaves your control except to the model API you pick. Powerful, but you'll want a human reviewing what it ships.

Type: Autonomous · self-host · Open source: Yes (MIT) · Models: Any (BYO key) · Your code: Self-hosted / your API · Pricing: Free + model cost · visit ↗

Aider · 3.7/5

A lightweight, git-native open-source pair-programmer in the terminal — brilliant for focused, well-scoped changes. Bring your own key, so code goes only to your chosen model; minimal setup.

Type: CLI · Open source: Yes (Apache-2.0) · Models: Any (BYO key) · Your code: Only to your model API · Pricing: Free + model cost · visit ↗

Kimi (Moonshot) · 3.5/5

Long context window and agentic coding at aggressive pricing — the value pick from the open-weight camp. Code is sent to Moonshot's hosted service, though the weights are open if you'd rather self-host. Worth a look if cost is the deciding factor.

Type: Web · agent · Open source: Weights open · Models: Kimi · Your code: To Moonshot (or self-host) · Pricing: Usage · visit ↗

How we score. Scored 0–5 on autonomy and reliability, model quality, how it fits a small team's workflow, and value. "Your code" describes where your source is sent at run time — to the tool's vendor, or only to whichever model API you point it at. Scores are RunAgentRun's editorial view; the factual fields are re-checked daily. Scores are RunAgentRun's editorial view, not the vendors'. Some outbound links may be affiliate — see our disclaimer.

Sources for the factual fields: Anthropic — Claude Code data usage · Cursor — privacy & Privacy Mode