Which AI agent should you choose? 5 questions, 3 picks

How to use it

Answer the five questions: the recommendation updates on every click, with no submit button and nothing sent anywhere. The tool searches the 388 agents in the directory (347 open-source projects and 41 commercial services) and keeps the three that best fit your situation. Each card gives the cost, the risk level, the reason for the pick and three links: the full profile, a guide to install it or get started, and a comparison when one exists.

• Technical level: a beginner won't see tools that only work from the command line or as a code library; a developer sees agents that live in the terminal or editor first.
• Task: it picks the family of agents (coding, personal assistant, automation, browser, computer use, research, agent building or business).
• Budget: a paid service whose entry plan exceeds your budget is dropped. At $0, only services with a free plan and free software remain.
• Hardware: it tells whether you can install software, run a local model or host a service.
• Confidentiality: with sensitive data, the tool requires a self-hostable agent that can run with a local model, so nothing leaves your machine.

How agents are ranked

The starting point is the directory score, out of 100, computed for each open-source project from public signals: popularity, recent activity, licence, documentation. Commercial services, which have no public repository, start from a neutral 60. The tool then adds or removes points based on your answers, with fixed rules that are the same for everyone:

• beginner: −30 for a tool with no graphical interface, −40 for a framework or infrastructure component, +10 for a ready-to-use service;
• browser only: −40 for software you must install; GPU or large Mac: +15 if the agent accepts a local model; server: +10 if it is self-hostable;
• sensitive data: −50 if the agent cannot run entirely on your side; another −15 if its risk is high;
• maintenance: −20 with no commit for six months, −40 for a year; high risk: −10, or −25 for a beginner.

Paid services whose entry price exceeds your budget are excluded, not just penalised. If fewer than three agents of the requested family pass the filters, the tool widens to neighbouring families with a penalty, and says so. No vendor pays to appear or rank higher.

Reading a recommendation

The cost of open-source software is zero, but you pay for the model it calls, unless you plug in a local model: 85 agents in the directory allow it. To estimate the API bill, use the Lab's price calculator. For a commercial service, the price shown is the vendor's published entry plan; team or enterprise plans cost more.

The risk follows the directory method: it rises when the agent can run commands, drive the computer or pay without default human approval, when it reads outside content that may carry booby-trapped instructions, when it sends messages, or when the project is no longer maintained. 77 agents are rated high risk: they are not to be avoided, but installed with guardrails. The guide Secure an AI agent explains which ones.

Before adopting an agent, run a short trial:

1. install it on a test account or folder, without your real credentials;
2. give it three tasks you know how to do yourself, so you can judge the result;
3. read its action log and note what it cost;
4. compare with the second and third in the list before deciding.

Where the data comes from

Each agent in the directory has a profile collected on October 25, 2026: type, interfaces (web, terminal, editor, messaging, app), supported models, self-hosting option, access granted, licence and last commit date for open-source projects, published entry plan for commercial services. The tool holds no hand-added data: it reads these profiles, embedded in the page, and computes in your browser. Your answers are neither sent nor stored.

Limits of the tool

• Five questions don't say everything. Integration with your tools (CRM, team chat, storage), interface language or customer support also matter: check them on the profile.
• Prices change. Entry plans are those published on the collection date; some are yearly, others billed by usage.
• The score is not a test. It measures a project's vitality, not the quality of its answers on your task. Only your own trial will tell.
• The rules are editorial choices. The points added or removed reflect our judgement, set out above so you can challenge it.