Skip to content
FrameworkHunt

Method

A directory is only worth reading if you can check it. This page describes exactly how a score is produced, what it is allowed to claim, and what makes the build fail.

Two layers, never blended

Each framework carries a measured layer and an assessed layer, and they are kept apart on purpose.

Measured facts come from the GitHub API: licence, language, commit recency, releases, contributors, issue cohorts, stars. They are dated, reproducible, and displayed without judgement. They are also never scored. A popular repository is not necessarily well designed and a quiet one is not necessarily unhealthy. The scoring function takes axis values as its only argument and throws if a measured key is passed to it — that rule is enforced by a test, not by convention.

Six axes

Each axis is scored 0–4 and weighted. The total is computed, never authored by hand; if a stored total disagrees with a fresh computation, the build fails.

Control flow & multi-agent composition20%
Sequences, routers, graphs, parallel work, loops, handoffs, and whether control is deterministic or model-routed.
State, memory & durability20%
Typed state, session vs long-term memory, checkpoints, persistence, resume/restart, retry, and idempotency semantics.
Tools, integrations & interoperability15%
Typed tool contracts, structured output, streaming, provider/plugin adapters, and native or official protocol support.
Reliability, safety & human controls15%
Timeouts, retries, cancellation, guardrail enforcement, permissions, isolation, approvals, interrupts, and secret boundaries.
Observability & evaluation15%
Run/trace identity, model/tool/handoff events, logs, metrics, cost visibility, replay, evals, and test fixtures.
Developer experience, portability & maintainability15%
Clean install, canonical quickstart, public API/types, supported runtimes, deployment path, versioning, and migrations.

What each level means

0
Absent — a scoped, reproducible search found no support.
1
Claimed — first-party mention exists, but no executable path is evidenced.
2
Basic — documented API or source shape exists, but behavior is manual, partial, or lacks durable semantics.
3
First-class — documented path plus a runnable official example or test demonstrates normal behavior.
4
Production-grade — level 3 plus evidenced failure, recovery, enforcement, or integration-boundary semantics.

Unknown is not zero. If the evidence is insufficient, the axis stays Unknown and the record publishes no aggregate score at all. A missing answer is never rounded down into a bad one.

The evidence contract

Every axis score, strength, limitation, recommendation and pattern references citations. A citation names a file path and the literal strings that must appear in it. A verifier fetches that file at the reviewed commit, and if the text is not there the build fails. It is not possible to publish a claim whose source does not contain it.

What this does not tell you

The score measures architecture as evidenced in a repository at one commit. It is not a benchmark: nothing here has been run against a workload, and no latency, cost or success-rate claim is made. It does not measure community support quality, commercial terms, or how a framework will feel on your problem. A framework scoring lower may be the right choice for you — read the axis columns and the “avoid if” lines rather than the total.

Reviews are point-in-time. Fast-moving projects change; the reviewed commit is printed on every page so you can see exactly how stale an assessment is.

Independence

No framework has paid to appear here, no listing is sponsored, and there are no affiliate links. Nobody has been given advance sight of their score.

Rubric v1.0.0 · 4 frameworks · 79 verified citations · citations verified 18 Aug 2026 · snapshot 18 Aug 2026