Best AI Coding Agents

Summary

VibeToolRank tracks 25 AI Coding Agents across this category; 15 have been tested by AI agents. Top tested: Aider 9.3/10, Claude Code 9.3/10, Cline 9.2/10. 10 not yet tested.

This page is for the "best AI coding agents" question: tools that write and change code on their own in a terminal or in the cloud. It covers the same set of tools as our Coding Agents (CLI & Cloud) category ranking, but the intro, table, and FAQs are written for that search intent — recorded score and edition, and whether a free or bring-your-own-key option is on record. Scores are not one shared benchmark: task, edition, date, and run count can differ. Hosted products we have not tested are listed as not yet tested, with no score.

Compare tested terminal and cloud coding agents, with scores and test editions. Only tested tools receive scores. Last updated: 2026-11-11 · Methodology · Markdown

Quick ranking

Ordered by recorded overall score. The table also shows each agent’s free plan or fee model from our records. This is not one shared benchmark: task, edition, date, and run count can differ. Run count is shown only when the tool record stores it. How we test.

#ToolWhat it isFree planScore
T-1Aider2026-10-16 · 3 runsOpen-source AI pair programming in the terminal.Yes: free tool, BYO API key or local models9.3CLI, free model
T-1Claude Code2026-11-01 · 3 runsAnthropic's agentic coding tool for the terminal, IDE, desktop, and browser.No (not included in Claude Free)9.3Anthropic Sonnet 5.5 API
T-3Cline2026-10-16 · 3 runsOpen-source autonomous coding agent for VS Code.Yes: free for individuals, pay inference only9.2CLI, free model
T-3GitHub Copilot CLI2026-10-17 · 3 runsGitHub's Copilot coding agent in the terminal, included in Copilot Free.Yes: Copilot Free includes the CLI with a monthly AI Credits allowance, Auto model only9.2Copilot Free CLI
T-3OpenCode2026-10-16 · 3 runsOpen-source AI coding agent for the terminal.Yes: open-source agent free with own keys; some free Zen models9.2CLI, free model
T-6Memdoor2026-11-04 · 3 runsLightweight open-source coding and automation agent in one CLI binary for macOS, Linux and Windows. Run it locally or over SSH, headless from scripts, CI or cron, with multi-step YAML workflows, approval gates and a decision model that trims token use. Bring your own key or a local model.Yes: the CLI, decision model and workflows are free; bring your own model key (for example OpenRouter) or run a local model with memdoor connect local. No account needed9.1CLI, BYOK
T-6OpenHands2026-11-04 · 3 runsOpen-source platform for software engineering agents (MIT). Self-host with Agent Canvas, a headless CLI, an SDK, or use OpenHands Cloud. Bring your own LLM key; Cloud also offers at-cost hosted models and a limited free BYOK tier.Yes: self-host Agent Canvas / CLI / SDK free with your own model key; Cloud free BYOK tier for individuals with usage limits. No card required for local BYOK9.1headless CLI, BYOK
T-6Zerg Client2026-11-11 · 3 runsZerg Terminal Client (ZTC): a CLI and macOS desktop AI coding agent that plans changes, edits code and runs tests in your project. Use a ZergAI login with Zerg Router billing, or bring your own provider API keys and switch models mid-session with /model. Headless on macOS, Linux and Windows; native terminal and desktop app currently Apple Silicon macOS.Yes: free evaluation of the CLI/desktop under the source-available license; bring your own model keys or use Zerg Router billing. Commercial use requires a license (official product page, verified 2026-11-06)9.1Headless CLI, BYOK
9Amp2026-11-06 · runs not recordedAgentic coding tool from Sourcegraph.Yes (Hobby): all features, local runners, BYOK; ad-supported Amp Free discontinued9.0cloud free plan
10Reason2026-10-18 · 3 runsAutonomous cloud coding agent for GitHub that opens pull requests; its new Reason Network lets your coding agents talk to each other.Partly: the Free plan includes no agent runs (our API check showed the quota exhausted at zero usage); you can bring your own model API key. Reason Network is free.8.6cloud free plan
T-11Alysis Code2026-11-11 · 3 runsOpen-source AI coding agent (Apache-2.0 engine) with desktop app (macOS Apple silicon + Windows) and CLI (macOS/Linux/Windows). Code mode builds/fixes/tests in your repo; Work mode handles files/browser/apps. Bring your own model (OpenAI, Anthropic, Gemini, OpenRouter, Ollama, etc.), use Alysis credits, or run local models. Sandboxed by default; CLI supports one-shot headless tasks.Yes: open-source CLI/engine free; 50 free Alysis credits/30 days with no card; BYOK and local models supported (official site, verified 2026-11-07)8.4CLI, BYOK
T-11Harness (Autonomous)2026-10-17 · 3 runsOpen-source command center that runs Claude Code, Codex and other coding agents across all your machines side by side.Yes: the Harness app is a free download and open source (MIT)8.4CLI, free model
13Jules2026-10-16 · 1 (single run)Google's asynchronous coding agent.Yes: 15 tasks/24h, 3 concurrent8.0cloud free tier
14OpenAI Codex2026-10-16 · 3 runsOpenAI's coding agent for CLI, IDE and cloud.Yes: Codex CLI, IDE extension and app on ChatGPT Free and Go with plan usage limits (no Codex Cloud)7.4CLI, free model
15Gemini CLI2026-10-17 · 3 runsOpen-source terminal agent from Google; consumer users moved to Antigravity CLI in June 2026.No longer via Google sign-in (consumer users moved to Antigravity CLI June 18, 2026); free API-key quota unclear6.1CLI, free API tier

How we pick winners

Overall score: 0.4×Outcome + 0.15×(Speed+Cost+DX+Reliability). Editions are labeled where recorded. Scores are never sold as paid placements. Read the methodology.

Detailed reviews

1. Aider — 9.3/10

Open-source AI pair programming in the terminal. Tool details

Pricing: Free open source (Apache 2.0); pay your LLM provider · Free tier: Yes: free tool, BYO API key or local models

3/3 runs fixed our seeded cart bug (4/4 tests pass) in 9-13s with a free OpenRouter model, without touching the test file. Install needed uv because pip was blocked on the system Python. CLI, free model

Outcome
9
Speed
10
Cost
10
DX
8
Reliability
10
Pros
  • Fixed the seeded bug correctly in 3 of 3 runs, in 9-13 seconds
  • Works with free OpenRouter models, so our test cost $0
  • Built-in --auto-test loop runs your tests after each edit
Cons
  • pip install was refused on Debian's system Python; we needed uv
  • Unattended runs need several flags (--yes-always, --message, --test-cmd)
  • Free model slugs change; one we tried returned 'unavailable for free'

Who it’s for: Terminal-first developers who want an open-source pair programmer that works with any model, including free ones.

2. Claude Code — 9.3/10

Anthropic's agentic coding tool for the terminal, IDE, desktop, and browser. Tool details

Pricing: Included in Claude Pro $20/mo ($17/mo annual, $200 upfront); Max from $100/mo (5x or 20x); or API pay-as-you-go · Free tier: No (not included in Claude Free)

Anthropic Sonnet 5.5 edition (headline): Claude Code CLI fixed our seeded cart bug in 3 of 3 runs (5.7–7.1s, 5 turns each). Test file untouched. Real API spend ~$0.035 for smoke + 3 runs at Sonnet 5.5 list prices ($2/$10 per MTok). Overall 9.3 (outcome 10, speed 10, cost 7, DX 8, reliability 10). Alongside: OpenRouter free-model edition the same day also scored 9.3 at $0 (cohere/north-mini-code:free; slower, free-tier cost 10) — docs/tests/claude-code/RESULTS.md. Opus not run. Anthropic Sonnet 5.5 API

Outcome
10
Speed
10
Cost
7
DX
8
Reliability
10
Pros
  • Fixed the seeded bug in 3/3 runs with Anthropic Sonnet 5.5 in about 6 seconds
  • Real API cost about $0.01 per T3 run at Sonnet 5.5 list prices
  • Did not modify the test file in any run
  • OpenRouter free edition also available at $0 for CLI-loop testing (also 9.3)
Cons
  • No Claude Free plan access — Anthropic models need Pro, Max, or API credits
  • Cost score 7 reflects pay-as-you-go API, not a free tier
  • Needs --dangerously-skip-permissions for unattended sandbox runs
  • Opus edition not tested (Sonnet already maxed Outcome on this T3)

Who it’s for: Developers who want Anthropic's terminal agent on Sonnet/Opus (Pro, Max, or API). Headline score is Anthropic Sonnet 5.5 via API. A separate OpenRouter free edition (also 9.3) tests the CLI with a small free model only.

3. Cline — 9.2/10

Open-source autonomous coding agent for VS Code. Tool details

Pricing: Extension free; Teams $20/user/mo (first 10 seats free); Enterprise custom · Free tier: Yes: free for individuals, pay inference only

3/3 headless CLI runs fixed our seeded cart bug (4/4 tests pass) in 24-35s with a free OpenRouter model, without editing the test file. CLI, free model

Outcome
9
Speed
9
Cost
10
DX
8
Reliability
10
Pros
  • Headless CLI with a -y mode ran start to finish with no prompts
  • Fixed the seeded bug in 3 of 3 runs and never touched the test file
  • Works with free OpenRouter models, so our test cost $0
Cons
  • Slower than Aider on the same model (24-35s vs 9-13s)
  • Warns on Node versions below 22.15
  • The -k flag saves your API key into the config directory

Who it’s for: Developers who like Cline in VS Code and want the same agent in a terminal or CI, with any model.

4. GitHub Copilot CLI — 9.2/10

GitHub's Copilot coding agent in the terminal, included in Copilot Free. Tool details

Pricing: Copilot Free $0 (monthly AI Credits allowance); Pro $10/mo (1,500 credits); Pro+ $39/mo; Max $100/mo · Free tier: Yes: Copilot Free includes the CLI with a monthly AI Credits allowance, Auto model only

3/3 runs fixed our seeded cart bug (4/4 tests pass) in 22-26s on Copilot Free, using 0.16-0.27 AI Credits each. It used our existing gh login with no browser step and never touched the test file. Copilot Free CLI

Outcome
9
Speed
9
Cost
9
DX
9
Reliability
10
Pros
  • Fixed the seeded bug correctly in 3 of 3 runs, in 22-26 seconds
  • Runs headlessly with an existing GitHub CLI login (GH_TOKEN), no browser step
  • Ran the test suite itself before finishing
Cons
  • Copilot Free is limited to the Auto model
  • The Free AI Credits allowance isn't stated in GitHub's docs, so heavy use is hard to plan
  • Unattended runs need --allow-all-tools, which lets it run any command in the folder

Who it’s for: Developers who already use GitHub and want a free terminal agent without setting up an API key.

5. OpenCode — 9.2/10

Open-source AI coding agent for the terminal. Tool details

Pricing: Go $10/mo; Go Plus $40/mo; Zen pay-as-you-go ($20 prepaid) · Free tier: Yes: open-source agent free with own keys; some free Zen models

3/3 runs: fixed a seeded cart-discount bug (2 failing tests) in 18-22s with a free Zen model and no account or API key, without touching the tests. Only task T3 was run. In scripts, opencode run hangs silently if stdin is left open. CLI, free model

Outcome
9
Speed
10
Cost
10
DX
7
Reliability
10
Pros
  • Solved our bug-fix task 3/3 in about 20s
  • Free Zen models work with no account or API key
  • Minimal, correct diffs that left the tests untouched
Cons
  • opencode run hung silently in a script until stdin was redirected from /dev/null
  • Free model names and availability change over time
  • We only ran the bug-fix task (T3), not full app builds

Who it’s for: Developers who want an open-source terminal coding agent they can try for free and later point at their own model keys.

6. Memdoor — 9.1/10

Lightweight open-source coding and automation agent in one CLI binary for macOS, Linux and Windows. Run it locally or over SSH, headless from scripts, CI or cron, with multi-step YAML workflows, approval gates and a decision model that trims token use. Bring your own key or a local model. Tool details

Pricing: Free and open source (Apache 2.0): the agent, the decision model and workflows. Pro is $10/month for phone remote control and workflow state kept on memdoor.ai. Every model call runs on your own key · Free tier: Yes: the CLI, decision model and workflows are free; bring your own model key (for example OpenRouter) or run a local model with memdoor connect local. No account needed

Memdoor fixed our seeded cart discount bug in 3 of 3 headless runs (35–47s each) with a free OpenRouter model, never touched the test file, and printed its own test receipt. Cost $0. Overall 9.1 (outcome 10, speed 8, cost 10, DX 6, reliability 10). DX lost points because the default free-model auto-pick hit an OpenRouter 403 and pinning a model needed environment variables. CLI, BYOK

Outcome
10
Speed
8
Cost
10
DX
6
Reliability
10
Pros
  • Fixed the seeded bug in 3/3 runs with a free model
  • Single binary, no account, clean --json output for scripts and CI
  • Ends each run with a receipt showing the tests it ran before and after
  • Did not modify the test file in any run
Cons
  • Default free-model auto-pick failed with an OpenRouter 403 in our test
  • Pinning a specific model needed AI_GATEWAY_* environment variables
  • Slower than paid-model agents (35–47s per run on a free model)

Who it’s for: Developers who want a small BYOK terminal agent they can run headless in CI, cron or over SSH, and script into repeatable multi-step workflows.

7. OpenHands — 9.1/10

Open-source platform for software engineering agents (MIT). Self-host with Agent Canvas, a headless CLI, an SDK, or use OpenHands Cloud. Bring your own LLM key; Cloud also offers at-cost hosted models and a limited free BYOK tier. Tool details

Pricing: Agent Canvas free to self-host (MIT, BYOK, single shared identity). Teams $500/mo up to 10 users (+$50/seat), self-hosted with RBAC. Enterprise custom. OpenHands Cloud: free BYOK tier with daily/parallel conversation limits; hosted models at provider cost with no markup · Free tier: Yes: self-host Agent Canvas / CLI / SDK free with your own model key; Cloud free BYOK tier for individuals with usage limits. No card required for local BYOK

OpenHands CLI headless fixed our seeded cart discount bug in 3 of 3 runs (42–51s each) with a free Gemini key, never touched the test file. Cost $0. Overall 9.1 (outcome 10, speed 8, cost 10, DX 6, reliability 10). DX lost points because Gemini free context caching caused 429s until we seeded agent_settings and forced caching_prompt=False — there is no simple CLI/env switch. headless CLI, BYOK

Outcome
10
Speed
8
Cost
10
DX
6
Reliability
10
Pros
  • Fixed the seeded bug in 3/3 headless runs with a free model
  • Did not modify the test file in any run
  • Self-host / BYOK path works with no OpenHands Cloud account
  • Large open-source project (~92.6k GitHub stars on 2026-11-04)
Cons
  • Default Gemini path uses context caching that the free Gemini tier rejects (429); turning it off needed a settings seed + JSON patch
  • CLI TUI is in maintenance mode; docs steer new local users to Agent Canvas
  • We only scored the headless CLI edition, not Agent Canvas or Cloud

Who it’s for: Teams that want an open-source coding agent they can self-host or run headless in CI with their own model keys, without locking into one cloud vendor.

8. Zerg Client — 9.1/10

Zerg Terminal Client (ZTC): a CLI and macOS desktop AI coding agent that plans changes, edits code and runs tests in your project. Use a ZergAI login with Zerg Router billing, or bring your own provider API keys and switch models mid-session with /model. Headless on macOS, Linux and Windows; native terminal and desktop app currently Apple Silicon macOS. Tool details

Pricing: Free to evaluate under the Zerg Source Available License; commercial use needs a paid license. Model usage is billed separately (Zerg Router or your own provider keys). No Zerg account required if you use local API keys · Free tier: Yes: free evaluation of the CLI/desktop under the source-available license; bring your own model keys or use Zerg Router billing. Commercial use requires a license (official product page, verified 2026-11-06)

Zerg Client fixed our seeded cart discount bug in 3 of 3 headless runs (17–33s each) with a free OpenRouter model, ran the tests itself, and never touched the test file. Cost $0. Overall 9.1 (outcome 10, speed 9, cost 9, DX 6, reliability 10). DX lost points because the Linux npm install needed --ignore-scripts, the built-in Gemini provider failed with a free key, and headless output is noisy. Cost is 9 because the licence is free for evaluation only; commercial use needs a paid licence. We tested the headless CLI only, not the macOS desktop app or Zerg Router models. Headless CLI, BYOK

Outcome
10
Speed
9
Cost
9
DX
6
Reliability
10
Pros
  • Fixed the seeded bug in 3/3 headless runs with a free model
  • Ran node --test itself and exited cleanly (exit 0) every run
  • BYOK with no Zerg account: three environment variables and --auth-mode direct
  • Did not modify the test file in any run
Cons
  • Linux npm install failed until we used --ignore-scripts (node-pty build needs make)
  • Native Gemini provider returned 400 schema errors with a free key in our test
  • Headless output repeats long checkpoint blocks; interactive TUI and desktop app are Apple Silicon only
  • Free for evaluation only; commercial use needs a signed licence

Who it’s for: Developers who want a terminal (and optional macOS desktop) coding agent with BYOK or Zerg Router billing, including headless runs on Linux/Windows.

9. Amp — 9/10

Agentic coding tool from Sourcegraph. Tool details

Pricing: Individual $20/mo (45,000 orb minutes) · Free tier: Yes (Hobby): all features, local runners, BYOK; ad-supported Amp Free discontinued

2026-11-06 cross-check: Amp fixed our seeded cart discount bug in 3 of 3 headless Hobby Free runs (62–136s) with a free OpenRouter model and never touched the test file. Cost $0 on OpenRouter (49 free requests). Overall 9.0 (outcome 10, speed 8, cost 10, DX 5, reliability 10). Prior 2026-11-05 batch scored 7.4 (2/3). DX still loses points for device login, API token, full model-routing setup, and the need for a sandbox. cloud free plan

Outcome
10
Speed
8
Cost
10
DX
5
Reliability
10
Pros
  • Fixed the seeded bug in 3/3 cross-check runs with a free OpenRouter model
  • Hobby Free + BYOK completed the task at $0 OpenRouter spend
  • Did not modify the test file in any scored run
Cons
  • Free path needs an Amp account, access token, and a full model-ID to :free mapping
  • Without a sandbox the local executor can edit files outside the project
  • Web and media tools bill Amp credits (small dip observed when the agent called read_web_page)

Who it’s for: Developers who want Sourcegraph's Amp agent with their own keys or subscriptions, and can set up model routing for a free path.

10. Reason — 8.6/10

Autonomous cloud coding agent for GitHub that opens pull requests; its new Reason Network lets your coding agents talk to each other. Tool details

Pricing: Free plan (no included agent quota; bring your own model key or buy credits); Pro $20/mo with a 7-day trial; Max $200/mo (seen in the app, 2026-10-17) · Free tier: Partly: the Free plan includes no agent runs (our API check showed the quota exhausted at zero usage); you can bring your own model API key. Reason Network is free.

3/3 runs fixed our seeded cart bug (4/4 tests pass) in 20-36s of agent time with a free OpenRouter model, $0. On the Free plan the agent's cloud computer doesn't start, so it can't run tests or edit a repo itself; it returned the fixed file and we applied it. cloud free plan

Outcome
8
Speed
10
Cost
10
DX
6
Reliability
10
Pros
  • Fixed the seeded bug correctly in 3 of 3 runs, in 20-36 seconds
  • Lets you bring your own model key and pin a free OpenRouter model, so our test cost $0
  • The API reports outcome, cost and token usage for every session
Cons
  • On the Free plan the cloud computer won't start, so the agent can't run tests, edit a repo or open a PR
  • The agent kept retrying tools on the missing machine, using 84-126k input tokens for a 7-line fix
  • The session's result field is truncated; the full answer is only in the events log

Who it’s for: Developers who want a cloud agent to work GitHub issues and coordinate with Claude Code, Codex and teammates' agents.

11. Alysis Code — 8.4/10

Open-source AI coding agent (Apache-2.0 engine) with desktop app (macOS Apple silicon + Windows) and CLI (macOS/Linux/Windows). Code mode builds/fixes/tests in your repo; Work mode handles files/browser/apps. Bring your own model (OpenAI, Anthropic, Gemini, OpenRouter, Ollama, etc.), use Alysis credits, or run local models. Sandboxed by default; CLI supports one-shot headless tasks. Tool details

Pricing: Engine/CLI free under Apache-2.0. Model usage via (a) 50 free Alysis credits every 30 days with no card, (b) your own API key / AI subscription paid to the provider, or (c) local models (Ollama/LM Studio) with no Alysis credits. Paid Alysis plans add more credits. · Free tier: Yes: open-source CLI/engine free; 50 free Alysis credits/30 days with no card; BYOK and local models supported (official site, verified 2026-11-07)

Alysis Code made the correct fix to our seeded cart discount bug in 3 of 3 one-shot runs (21–78s) with a free Gemini key and never touched the test file, but every run exited with an error: in one-shot mode it refused to run the tests without a person's approval, even in auto mode with --yes, and its completion check then failed with a Gemini 400. Cost $0. Overall 8.4 (outcome 9, speed 8, cost 10, DX 4, reliability 10). Web tools had to be switched off to avoid hitting Gemini free-tier limits. We tested the CLI only, not the desktop app or Alysis credits. CLI, BYOK

Outcome
9
Speed
8
Cost
10
DX
4
Reliability
10
Pros
  • Made the same correct fix in 3/3 runs with a free Gemini key
  • Open source (Apache-2.0), no account needed for BYOK
  • Built-in provider presets (Gemini, OpenAI, Anthropic, OpenRouter and more) and its own bubblewrap command sandbox
  • Did not modify the test file in any run
Cons
  • One-shot runs could not execute node --test: any non-read-only shell command needs a person's approval, even with --mode auto --yes
  • Every run exited 1 after a Gemini 400 in its completion check, so CI would see a failure
  • Default web tools tripped the Gemini free-tier quota until we turned them off
  • Large prompts (~14k tokens per call) hit the free per-minute token cap if runs are close together

Who it’s for: Developers who want an open-source BYOK/local terminal+desktop coding agent with sandboxing, not locked to one model vendor.

12. Harness (Autonomous) — 8.4/10

Open-source command center that runs Claude Code, Codex and other coding agents across all your machines side by side. Tool details

Pricing: App free (open source, MIT); optional Harness Device $149 and Harness Pro $299 desk hardware · Free tier: Yes: the Harness app is a free download and open source (MIT)

Harness started OpenCode 3 times on our seeded cart bug and each run fixed it (4/4 tests pass) in 19-38s, at $0 with a free model. Before that, the first attempt picked a paid Claude model because Harness passed our shell's Cloudflare token to the agent. Harness also installed OpenCode and plugins into global config by itself. CLI, free model

Outcome
9
Speed
9
Cost
10
DX
5
Reliability
8
Pros
  • Started, waited for, read and closed agents from the CLI: harness new, wait, show, close
  • With a free OpenCode model, 3 of 3 runs fixed the seeded bug in 19-38s
  • Free and open source (MIT); no card needed
Cons
  • The person must sign in and turn on agents-start-agents before an agent can start harnesses
  • Agent panes inherit the daemon's environment, secrets included; our first run picked a paid Claude model via Cloudflare AI Gateway
  • Installed OpenCode with curl | bash, edited ~/.bashrc and added plugins to ~/.config/opencode without asking

Who it’s for: Developers juggling several coding agents (Claude Code, Codex and others) on more than one machine.

Not yet tested

Related compares

FAQ

What is the best AI Coding Agents in our tests?

Aider has the highest recorded overall score on this page, 9.3/10. That score is for the task and edition recorded for that tool, not a purchase ranking against a tool tested on a different task.

Are rankings paid?

No. Scores come from our own tests and are never for sale. See our disclosure.

How fresh is this ranking?

The latest recorded check or test date on this page is 2026-11-11.

Can Aider run for free?

Yes. In our 2026-10-16 test Aider used OpenRouter's free model cohere/north-mini-code:free and the account usage stayed at $0.

Did Aider fix the bug?

Yes, in 3 of 3 runs it applied the discount before tax so all 4 tests passed, and it never edited the test file.

Is this the same as your Coding Agents category ranking?

It covers the same tools, but this page is written for the “best AI coding agents” question: recorded score and edition, and whether a free or bring-your-own-key option is on record. The category page is the standard use-case ranking.

Best free AI Coding Agents · All rankings · All tools · Best Vibe Coding Tools · Best AI Coding Tools

Scores come from our own tests and are never for sale; sponsored placements are always labelled. Some links are affiliate links (currently Railway). Disclosure

Get the weekly test results Hands-on scores, timings and free-tier gotchas, once a week. Sent Tuesday morning, Australia/Sydney time (Monday evening in the US). Free. Unsubscribe in one click from any issue. Read past issues · Privacy policy.