Ecosystem Report: April 2026 — Benchmarks and Autoresearch

Key Takeaways

• +37 new projects added in April (vs +25 in March) — fastest growth month yet
• Hermes Agent broke out: +44,092 stars (+54%) — by far the top mover among projects
• The autoresearch cluster: a dense group of autonomous-research projects landed in mid-April, signalling concentrated builder attention on the category
• Benchmark surge: 7 new evaluation suites (Claw-Eval, ClawBench, ClawMark, BioAgent Bench, BixBench, ResearchClawBench, HeurekaBench)
• 17 projects officially marked stable (paper-locked, intentionally low-activity)
• First entry in the Graveyard: PaperClaw (meowscles69, 141 stars) lost to a GitHub account deletion
• Tracked total: 144 projects + 33 skill hubs = 177 entries (was 119 at end of March)

The Numbers

The 116-repo subset is what we have continuous daily snapshots for — full coverage starts mid-April when daily snapshots stabilised.

Top Movers — Projects (April 14 → April 30)

By absolute stars added:

By percentage (start ≥ 20 stars, to filter noise):

Hermes is the headline. The +44K story is real growth — it correlates with their April release cycle and the "agent harness" category becoming legible to the broader Claude / OpenClaw community. Among projects, nothing else came close in absolute terms; OpenHarness was a distant second.

For comparison, the largest skill-hub mover was anthropics/skills at +9,375 — meaningful, but a different category and not directly comparable to project repos.

The Autoresearch Cluster

A dense cluster of autonomous-research projects came onto our radar in the second half of April:

• SakanaAI/AI-Scientist — the original (v1)
• davebcn87/pi-autoresearch
• uditgoenka/autoresearch
• InternScience/InternAgent
• trevin-creator/autoresearch-mlx (Apple Silicon)
• jimmc414/Kosmos

We added them on the same day, which reflects when we found them, not when they launched. The signal is the concentration — six independent autoresearch projects surfacing in one curation pass means the category is getting active developer attention right now. Combined with the existing AutoResearchClaw, EvoScientist, NeuriCo, and DeepScientist, "autonomous scientific research" is now a 10+ project category.

Benchmark Surge

Seven new evaluation suites landed in April, signalling that science-agent evaluation has matured enough that people are building dedicated benchmarks instead of one-off tests:

• Future-House/BixBench (4/20) — the headline benchmark
• bioagent-bench/bioagent-bench (4/9) — bioinformatics-specific
• claw-eval/claw-eval (4/8)
• reacher-z/ClawBench (4/14)
• evolvent-ai/clawmark (4/14)
• weibowen555/ClawSafety (4/7)
• mlbio-epfl/HeurekaBench (4/1)

The "benchmark proliferation" pattern usually precedes a "consolidation" phase where one or two dominate. Watch BixBench (Future House lab + arXiv paper) and BioAgent Bench (clear bio scope).

New Stable Tier

We introduced a status: "stable" flag in April for paper-locked projects — repos that are intentionally low-activity because the paper is the deliverable, not the codebase. 17 projects now carry this badge:

BixBench, ChemCrow, AI-Scientist (v1 + v2), Virtual Lab, MDCrow, BioDiscoveryAgent, CRISPR-GPT, Agent Laboratory, AutoBA, AI-Researcher, Auto-Deep-Research, freephdlabor, BioMaster, GAIR-NLP/SR-Scientist, OmniScientist, BioAgent Bench.

This stops the dormant-detector from raising false alarms and gives readers a clearer signal: "no commits ≠ abandoned" when the paper anchors the work.

First Tombstone: PaperClaw (meowscles69)

April also brought the first entry to our Graveyard. PaperClaw by meowscles69 (141 stars at death, 27 OpenClaw skills covering literature reviews, hypothesis versioning, grant writing, lab knowledge handoffs) disappeared on April 13 when the GitHub account was deleted.

It was the second-most-popular of the eight projects sharing the "PaperClaw" name. The deletion took the code, issues, and discussions with it. If you were using it, three sibling projects still exist:

• guhaohao0991/PaperClaw (205⭐) — paper search-review-critique
• thinson/RS-PaperClaw (49⭐) — remote sensing focus
• hurtjan/PaperClaw (3⭐) — PDF citation graph

The loss is a reminder of why we run weekly health checks — and why archiving a clear eulogy matters more than just removing the entry.

Notable Arrivals

• OpenHarness (HKUDS, 4/2) — generic agent runtime, +23% in two weeks
• NeuriCo (ChicagoHAI, 4/2) — autonomous research from U Chicago, paper-anchored
• DeepScientist (4/3) — fully autonomous scientific reasoning
• StatsClaw (4/4) — multi-agent disagreement via information isolation
• TCM-Agent (4/6) — traditional Chinese medicine, niche but active
• HealthClaw (4/8) — HC-Guo, clinical research focus
• Bioinfor-Claw (4/17) — went 7 → 56 stars in 13 days post-listing
• Yuan1z0825/nature-skills (added 4/30) — Nature-standard figures + polishing, 265⭐ in 6 days

What Went Quiet

April's add velocity slowed in the second half of the month. After the autoresearch cluster surfaced, only 4 entries arrived between 4/15 and 4/30 (Bioinfor-Claw, BixBench, two PaperClaws). Either the field is saturating short-term, or builders are heads-down on May releases.

What to Watch in May

1. Hermes Agent crossing 150K stars — current trajectory says mid-May
2. BixBench citations — will it become the standard bio-agent benchmark?
3. OpenHarness vs autoagent — agent harness category consolidation
4. qinheming/bioclaw — went 0 → 103 stars in April, on the watching list
5. First "I tested X" hands-on review post — long-promised, high-trust content

Methodology

• Star counts: GitHub stars via our /api/cron-refresh endpoint, snapshotted daily into docs/stars-history.csv
• Project count: live from src/config/locale/messages/en/pages/index.json (projects) and skill-hubs.json (hubs); growth comparisons separate the two so we don't put apples next to oranges
• Growth window: April 14 → April 30 (16 days). Earlier data not consistent enough for full-month comparisons; will be a true 30-day window from May onwards
• stable status: human-curated for paper-anchored projects with verified Nature / arXiv / journal publications

---

Previous report: Ecosystem Report: March Was Wild — 54 → 107 projects in March.

Last updated: 2026-05-03.