New tools, real work. Learn where to start, what to choose, how to set it up, and how to use it. Finish a first task, then build a routine that works.
Our editorial starting point
Try the same three real tasks in both. Keep the one that needs less correction. Start with the app’s default model; use deeper reasoning for a task that actually benefits from it. You do not need to chase every new benchmark leader.
A natural alternative if you prefer Google’s platform and connected apps. Check which integrations your account supports. Google apps
Choose it if you prefer the xAI/X ecosystem. Platform fit is a valid reason to choose an assistant; it is not a benchmark win. Grok overview
Artificial Analysis Intelligence Index v4.3.2
Checked · High, not Max/Xhigh
| Tested model / setting | AA index |
|---|---|
| Claude Opus 5.5Anthropic · high with fallback | 54 |
| GPT-6 AstraOpenAI · high | 51 |
| Grok 4.7xAI · high | 46 |
| Gemini 3.8 FlashGoogle · high | 41 |
Four selected models, not the full leaderboard. This snapshot is not live, a percentage, or an app rating. High does not mean equal compute across providers; app settings may differ. Scores do not measure usability, file access, or total cost. AA methodology
Web is enough for most work. Use the official desktop app for local files, or Codex / Claude Code if you are comfortable in a terminal. An alternative harness is optional.
Enough for most everyday work: questions, drafts, research, and uploaded files. No installation needed.
Browser access does not mean access to your local folders.
A practical home for files, long tasks, and reviewing outputs side by side.
Grant folder and computer-control permissions deliberately.
Capture ideas, use voice, and check a task while away from your desk.
Review consequential edits on a larger screen. Mobile features vary by provider.
Use when files, scripts, or an existing development workflow are central.
More control, more setup. Not a prerequisite for productive AI use.
Start here for writing, learning, research, and questions about documents. A chat app is usually enough.
Our pick: test both on the same three real tasks; keep the one whose output needs less correction. There is no universal winner.
Need to work directly on local files, but not comfortable in a terminal? Use the official desktop app with a supported local-project or folder connection.
Give access to a small working folder, not your whole drive. Verify feature availability on your OS and account before paying. Desktop does not mean all processing stays on your computer.
Comfortable in a terminal? Start with the standard Codex or Claude Code harness for project files, commands, and code changes. Otherwise, use its graphical interface.
Our advice: keep the standard harness if it does the job. Try access included in your existing plan before adding another client; compare correctness, review effort, and total cost. Back up before edits.
A reasonable free-first terminal option for coding and project work with an eligible personal Google account. No paid subscription is required for the individual free tier.
Weekly limits apply. Keep review prompts, use a separate working folder, and compare results on your own tasks before paying for more capacity.
For people who specifically want provider choice and are comfortable configuring a coding agent.
The client and inference bill are different things. Check supported authentication and provider terms; a chat subscription is not a universal API pass.
Consider it for X-centered research, live web discovery, and voice. Check the underlying source before using a claim in your work.
It is an assistant choice, not automatically a replacement for a local project agent. Buy it for a workflow you actually use.
Want to try a CLI without a paid plan? Antigravity CLI is a free-first Google option for eligible personal accounts, with weekly limits.
Set up Antigravity CLIStay with Codex or Claude Code when they meet your needs. Choose the open-source OpenCode client when you specifically want to switch providers or use open-weight models through OpenRouter, another API service, or a supported local endpoint.
The client, model, and hosting are separate choices. Open-source software does not make inference free; open weights do not make a hosted API local or private. Check billing, model licensing, tool support, and provider data handling.
Set up OpenCode + OpenRouterUSD reference prices checked on the date above, before tax. Region, app-store billing, promotions, and plan changes can differ. Confirm checkout; this is not a live price feed.
$0 / month
Eligible personal Google accounts, age 18+ in supported regions. Basic weekly limits; paid upgrades are optional. Google Workspace and Cloud access have separate eligibility and billing. Checked 2026-10-01.
$20 → from $100 / month
Pro offers higher usage tiers. Check the current tier price and account limits; do not assume unlimited runs. API billing is separate.
$20 → $100 / $200 / month
Monthly Pro; Max has 5x / 20x usage tiers. Claude Code is included in Pro and Max, subject to limits. Annual Pro is a different billing commitment.
Verify at checkout
The public plans page did not expose a verifiable price in this review. Check web vs app-store pricing, region, limits, and any X bundle separately.
Provider-dependent
Set a small provider budget or prepaid balance. Check whether your chosen login uses a supported subscription or metered API billing.
Seats + possible usage charges
For company-approved data use, administration, and contractual controls. Enterprise is a governance decision, not a promise of better answers. Compare the written offer.
Use one paid plan for a week. Record completed tasks, corrections, and time lost to limits. A $180 upgrade must recover more than $180 of useful time. At your own valuation of $30/hour, that is six hours a month—not a promise of savings.
Use an approved organizational plan when the work needs central access control, offboarding, auditability, or contractual data handling. Check feature entitlements, retention, minimum seats, and overage rules with procurement. Do not put confidential company files into a personal account while waiting.
This is a starting recommendation, not a performance ranking. Your inputs stay on this page; nothing is sent to a provider.
Your starting point
Choose one roughly $20/month plan after a small trial. Do not buy several overlapping subscriptions on day one.
Follow the setup stepsThe engine that generates answers. Start with the default; use deeper reasoning when a task genuinely needs it.
The workspace around the model: files, tools, memory, permissions, and task execution.
Your access, usage allowance, and administration. Paying more does not automatically make every answer better.
These are instructions to follow yourself. This page does not install software, change permissions, connect accounts, or purchase a subscription.
codex --sandbox read-only --ask-for-approval on-requestagy{
"$schema": "https://opencode.ai/config.json",
"permission": { "*": "ask" }
}Goal: Turn the attached notes into a one-page weekly update. Audience: My team. Inputs: Use only the files I supplied; label missing information. Output: A draft with decisions, owners, deadlines, and source references. Boundaries: Preserve originals. Do not send, publish, purchase, delete, or change permissions without asking. Check: Verify names, dates, totals, and every link. List anything I must review.
A sandbox limits where code can act. Approvals decide which actions may proceed. Neither makes a result correct, and a local sandbox does not stop an authorized cloud connector from sending a message.
Read-only for exploration. For edits, allow one working folder with a backup. Keep the tool’s sandbox enabled where supported and check what it actually restricts.
Pause for sending, publishing, spending, deletion, new integrations, secrets, or production changes. For repeatable low-risk work, use narrow allow rules—not an approval for every future command.
Do not grant unrestricted access just to remove interruptions. If an advanced workflow genuinely needs it, use a disposable isolated environment with no personal files or production credentials, restricted network access, and a rollback plan. A container with your home folder mounted is not enough.
A web page, email, repository, or document can contain malicious instructions. Do not let it authorize new tools, data uploads, or payments. Inspect unexpected permission requests instead of repeatedly approving them.
Checklist progress is session-only and resets on reload. Checking a box changes no app permissions.
These are workflow examples, not a guarantee that every app supports every connector. Check your provider’s directory, your plan, and your company’s policy. Prefer official integrations; an MCP server is software you are trusting.
One reference folder, read-only where supported.
Draft a brief with links back to the source documents.
Read availability or draft replies first; approve sends and invites.
Prepare tomorrow’s agenda without contacting anyone.
Named channels only, if the connector supports it.
Summarize decisions; do not post automatically.
A separate work profile with only necessary logins.
Gather sources. Approve purchases, submissions, and account changes.
One repository and a separate branch; no production secrets.
Make a small fix, run tests, and show a diff before publishing.
If the integration cannot limit access enough, upload selected copies instead. Never paste passwords or API secrets into a prompt. Check what data leaves your machine, revoke unused connections, and test write actions in a disposable destination first.
Start with two independent tasks. More agents can mean more cost, conflicting edits, and a longer review queue—not more finished work.
One thread per outcome. Put research in one thread and a draft from frozen inputs in another. Give each its own output folder.
One owner per document. For code, use separate branches and worktrees. Do not ask two agents to rewrite the same file.
Ask for changed files, checks, and unresolved questions. Integrate one result at a time. Stop starting tasks when review becomes the bottleneck.
Not for one assistant. Consider Herdr when you already run several terminal agents and need a shared view of working, blocked, and finished sessions. It organizes sessions; it does not choose the best model or include model credits.
After following its install guide, run herdr from a project directory, create a workspace per project, and launch codex, claude, or opencode in a pane. Keep each agent’s own permissions. Detaching or closing the terminal can leave agents running: stop the session explicitly when finished. A workspace is not a security sandbox.
A faster daily loop: write a five-line brief → let the agent work → review an artifact → save the useful instructions. Reuse a tested template or skill before adding another agent. Use a second assistant for a targeted critique, not an endless debate.
Optional references. Pick a category when your current setup needs something more.
Agents, open-weight models, and orchestration tools for getting work done. These are existing directory entries, not newly verified recommendations. Review current setup instructions, licenses, and permissions before installing; start in a separate workspace.
Self-hosted agent with persistent memory that writes its own reusable skills from experience, runs scheduled jobs, and connects to Telegram, Discord, Slack, and the terminal.
Local-first personal AI assistant that runs on your own devices, supports every major LLM, and connects to 20+ messaging channels. Crossed 100K GitHub stars within its first week (Jan 2026).
Extensible open agent framework from Block that automates engineering tasks end-to-end, built natively on MCP for tool connections.
Lets LLMs run code on your machine from natural language — a local, open counterpart to hosted code-execution sandboxes, with safeguards you control.
Platform for AI software development agents that write code, run commands, browse the web, and fix issues end-to-end. Formerly OpenDevin.
Terminal pair-programmer that edits code in your local git repo, works with Claude, GPT, and local models, and commits with sensible messages.
Open-source autonomous coding agent inside VS Code — creates and edits files, runs commands, uses the browser, with human approval at each step.
Open-source terminal coding agent in the Claude Code mold — provider-agnostic, scriptable, with a polished TUI. Run it on Claude, GPT, or open-weight models.
Open IDE assistant for VS Code and JetBrains — chat, autocomplete, and edit with any model, plus shareable team configs. The bring-your-own-model default for IDE AI.
Makes websites accessible to AI agents — extracts interactive elements and lets any LLM click, type, and navigate. The de-facto standard library for browser agents.
High-throughput LLM serving engine with PagedAttention — the default way to serve open-weight models (including Korean models like EXAONE and Solar) in production.
One-command local model runtime — pull and run open-weight models (EXAONE and HyperCLOVA X SEED included) on a laptop, with an OpenAI-compatible local API.
Serving engine with RadixAttention prefix caching and a structured-generation language — the main open rival to vLLM, strong on agentic workloads with repeated prefixes.
Unified gateway that speaks the OpenAI format to 100+ providers — one proxy for keys, budgets, fallbacks, and routing across Claude, GPT, Gemini, and Korean APIs alike.
LG's frontier Korean-English model family, including reasoning and multimodal variants, under one of Korea's five MSIT sovereign-AI mandates.
Naver's open-weight SEED lineup — efficient Korean-first models for deployment, from small on-device sizes to the SEED Think reasoning line.
Upstage's Solar open models — the first Korean family recognized as frontier-class by Artificial Analysis, with open releases alongside the commercial API.
Kakao's Kanana model family with open-weight releases — compute-efficient bilingual models built for KakaoTalk-scale consumer deployment.
Microsoft's multi-agent framework — agents that converse, critique, and delegate to each other, with the AgentChat API and no-code Studio on top.
Role-based agent teams — define a crew with roles, goals, and tools, and let them execute processes together. The fastest on-ramp from idea to working multi-agent flow.
Google's open framework behind Vertex AI agents — hierarchical agent composition, built-in evaluation, and deployment paths from laptop to cloud.
Agent framework from the Pydantic team — type-safe tools and structured outputs with the validation rigor Python backends already trust.
Build LLM-powered applications and graph-based agent workflows (LangGraph)