Skip to content
AgentHub

✎ Knowledge Base

How we score personal AI agents

Every ranking on AgentHub comes from four published axes and a fixed formula. Nothing is pay-to-play, and every input can be checked.

Principles

  • Transparent

    Weights, signals and formulas are public on this page.

  • Multi-axis

    No single number hides trade-offs. Sort by the axis you care about.

  • Reproducible

    Capability tests use a fixed task suite anyone can rerun.

The formula

Each agent gets a 0–100 score on four axes. The overall score is a weighted average, rounded to one decimal. Capability carries the most weight because it reflects what an agent actually does for you.

overall = 0.25 × Popularity
        + 0.35 × Capability
        + 0.20 × Control
        + 0.20 × Ease
Popularity 25%Capability 35%Control 20%Ease 20%

The four axes

Popularity

25%

How widely the agent is used and talked about right now.

  • GitHub stars and 30-day star velocity (open source)
  • Reported or estimated active users (hosted)
  • Search interest and community mentions

Capability

35%

What the agent can actually get done, measured on a fixed task suite.

  • Completion rate on the AgentHub personal-task suite
  • Breadth of integrations and tools
  • Memory, scheduling and multi-step reliability

Control

20%

How much ownership you keep over data, models and behavior.

  • Self-hosting and model choice
  • Data export, deletion and local storage
  • Permission granularity and audit logs

Ease

20%

How quickly a non-expert gets real value.

  • Time from sign-up / install to first completed task
  • Quality of onboarding and docs
  • Availability across devices and apps

Worked example

This week’s #1, OpenClaw, computed live from the same function the rankings use:

AxisScoreWeightContribution
Popularity97× 0.2524.25
Capability88× 0.3530.80
Control95× 0.219.00
Ease64× 0.212.80
Overall86.9

Data & cadence

  1. 1

    Collect

    Public signals (stars, releases, docs) are pulled automatically; usage figures come from vendors or third-party estimates, labeled as such.

  2. 2

    Test

    Each agent runs the Personal Task Suite in a clean account with default settings. Runs are recorded.

  3. 3

    Normalize

    Raw signals are converted to 0–100 with percentile scaling, so one outlier can’t flatten the field.

  4. 4

    Publish

    A new snapshot goes out every Monday. Previous snapshots stay online so movements are auditable.

How we stay neutral

AgentHub isn't affiliated with any agent listed. Agents are ranked neutrally, builders are promoted openly, and users always come first. These rules keep those apart:

  • Rankings are never for sale

    No one can pay to change a score, a rank or a review. Scores come only from the formula on this page.

  • Paid means labeled

    Anything paid carries a visible Sponsored, Featured or Affiliate label wherever it appears: featured builders, featured jobs, event sponsors and affiliate links.

  • Ranking code can't see the money

    The scoring module only receives agent scores. It doesn't import builder, job, sponsor or claim data, and a lint rule fails the build if anyone tries.

  • Mentors and sponsors have no say

    Incubator mentors and sponsors take no part in scoring. Mentors are recused from anything involving the teams they mentor.

  • Incubated teams get no head start

    Startups in the AgentHub Incubator are scored with the same public methodology and marked with an Incubator badge.

  • Claimed pages don't change scores

    Agent makers can claim their page to post updates and send corrections with evidence. Claiming never changes scores.

  • Our build partner is separate

    ReactiveSpace builds agents for clients, separately from our rankings. Its work and clients get no ranking favors.

Corrections

Think a score is wrong? Send evidence through the community. Accepted corrections are applied in the next snapshot and noted in the changelog. Agent makers can also claim their page to send corrections directly.

Glossary

Personal AI agent
Software that acts on your behalf across apps and time — not just answering, but doing tasks, remembering context and following up.
Self-hosted
You run the agent on your own machine or server. You choose the model and keep the data.
Hosted
The vendor runs the agent for you. Easier to start, with less control over data and models.
Skill / tool
A packaged capability an agent can call, such as sending email, booking a table or querying a database.
Long-term memory
Information an agent keeps between conversations, like preferences, people and ongoing projects.
Personal Task Suite
AgentHub’s fixed set of real-world errands used to measure Capability reproducibly.