✷ AI Open-Source Radar
Archify More Than Doubled and Took First from DeepSeek Harness
The week’s two biggest movers are both new to this ranking, and both roughly doubled in seven days: Archify took first place outright, and a prompt library for image generation took third. Between them sits DeepSeek Harness, still the largest project on the board and still second by absolute gain, but at well under half the rate it managed last week. Beneath the top three, the pattern that carried last week’s ranking has not gone away so much as split. Three of the top ten now sell agents rules rather than capability, and they disagree about whether a rule should forbid or instruct.
#1 tt-a1i/archify
+19,312 stars this week → 34,466 · Star curve
Earlier: 13,033 stars on 2026-08-16
Archify turns a codebase, or a written description of one, into an interactive map. The load-bearing word is validated: the map is rendered deterministically from JSON the agent emits, so a relationship either resolves against the source or it does not get drawn. That is a different bet from everything else in the top ten. The others constrain what the agent writes; Archify constrains what you are shown about what it wrote, and the failure it is built against is a confident diagram of a system that does not exist.
#2 deepseek-ai/deepseek-harness
+16,749 stars this week → 204,683 · Star curve
Last week: +42,534, ranked #1 · Earlier: 83,753 stars on 2026-08-14
Still the largest project here, and still second by stars added — but the drop from last week’s rate is the sharpest on the board. Its ecosystem is arriving on a separate line: a native desktop client for it entered our directory this week, at the bottom of this page.
#3 freestylefly/awesome-gpt-image-2
+12,944 stars this week → 25,629 · Star curve
Earlier: 9,052 stars on 2026-08-03
The odd one in the top three, and the only entry here not pointed at code. It converts prose image prompts into an atomic schema — subject, lighting, materials and layout as separate fields — and ships 544 reverse-engineered cases to fill it. Read it next to Ponytail and Skills below and it is the same idea in a different medium: stop asking the model nicely and hand it a protocol.
#4 DietrichGebert/ponytail
+8,369 stars this week → 117,159 · Star curve
Last week: +5,000, ranked #6 · Earlier: 94,481 stars on 2026-08-03
Ponytail is the forbid half of that argument. Its decision ladder makes an agent prove a feature is necessary, then prove that neither the standard library nor the platform already does it, before it is allowed to write anything new. It is one of only two returning names that gained more this week than last — and the other one barely moved.
#5 mattpocock/skills
+8,019 stars this week → 241,835 · Star curve
Last week: +14,911, ranked #2 · Earlier: 209,688 stars on 2026-08-08
Skills is the instruct half. TDD, domain modelling and debugging discipline, packaged as instructions you install into an agent, aimed at the same failure Ponytail attacks from the opposite side: where the ladder says don’t, these say here is how. It is the biggest project on this board by total stars, and its rate roughly halved from last week.
#6 stablyai/orca
+5,522 stars this week → 57,377 · Star curve
Last week: +5,420, ranked #5 · Earlier: 36,521 stars on 2026-08-03
Unchanged from last week in every respect that matters — same rate, one place lower. It runs several CLI coding agents in parallel isolated worktrees so you can compare their output before merging one.
#7 MadsLorentzen/ai-job-search
+5,326 stars this week → 38,512 · Star curve
Earlier: 29,368 stars on 2026-08-03
The board’s reminder that agent tooling is not all built for engineers: a Claude Code project that scrapes postings, judges fit, and produces tailored ATS-ready documents.
#8 K-Dense-AI/scientific-agent-skills
+5,016 stars this week → 39,193 · Star curve
Earlier: 32,971 stars on 2026-08-08
The instruct pattern again, moved out of software entirely: 163 standardized procedures across biology, chemistry, medicine and physics, so a general-purpose agent can run a research workflow without someone wiring up the databases and instruments first. Two entries built the same way, four ranks apart, written for completely different readers — that is the clearest sign this week that a skill has become a format rather than a product.
#9 openai/codex
+4,954 stars this week → 120,085 · Star curve
Last week: +8,915, ranked #4 · Earlier: 104,761 stars on 2026-08-08
A local coding agent from OpenAI that runs in your terminal. Its rate roughly halved from last week.
#10 calesthio/OpenMontage
+4,837 stars this week → 54,629 · Star curve
Earlier: 44,713 stars on 2026-08-03
An agentic video pipeline — research, script, assets, edit, render. It is not alone this week: the breakout board below carries a JSX-based video editor up more than threefold, and third place is an image project. Generative media is moving across all three boards.
Breaking out
Ranked by growth relative to their own size, not by stars added. These are small enough that a week can multiply them — and small enough to have escaped you.
sapientinc/PRAXIST · +1120.5% over 2.8 days → 4,528 stars · Star curve
Measured across less than three days, which is why the percentage is enormous and why it means less than it looks: two readings is not a shape. What it claims to be is an autonomous research system that turns runnable, measurable projects into continuous research runs and synthesises across generations of them. Worth a bookmark and another look next week, not a conclusion.
JetBrains/go-modern-guidelines · +257.2% this week → 2,968 stars · Star curve
Earlier: 810 stars on 2026-08-15
The instruct pattern once more, this time from JetBrains and aimed at a failure you can check: models write the Go they saw in training, not the Go that exists now. Guidelines plus tooling to force current idioms.
diffusionstudio/editor · +238.8% this week → 1,884 stars · Star curve
Earlier: 495 stars on 2026-08-16
A video editor whose document is JSX, so an agent can edit a timeline the way it edits code and read the result back. Bidirectional, which is the part that matters — most agent video tools can write a project file but cannot see what came out.
workweave/router · +219.1% this week → 3,092 stars · Star curve
Earlier: 953 stars on 2026-08-15
A drop-in proxy that scores models locally and routes each request to whichever is currently winning. Infrastructure built on the assumption that you will not commit to one model.
anthropics/claude-plugins-community · +208.4% this week → 2,874 stars · Star curve
Earlier: 348 stars on 2026-08-15
A vetted marketplace for plugins that extend Claude Code and Claude Cowork. Same shape as the skills entries above: the artefact worth noticing is the distribution format, not any single plugin.
New to the directory
Repositories we catalogued for the first time this week. Most are not new projects — we simply had not indexed them before.
- exo-explore/exo — clusters the machines you already own into one pool big enough for models that do not fit on any of them (47,144 stars)
- google-research/timesfm — Google Research’s decoder-only foundation model for time-series forecasting, now natively multivariate (28,314 stars)
- titanwings/distilly — distils a person’s voice and judgement out of their messages into a reusable agent profile (24,198 stars)
- agentsmd/agents.md — an open format for the AGENTS.md file that hands a coding agent your project’s context (23,999 stars)
- anywhere-labs/dsh-desktop — a native desktop client for DeepSeek Harness, the project ranked second above (22,082 stars)
That last one deserves a second look. The harness at number two is decelerating hard, and in the same week a 22,000-star desktop client for it turns up in our directory for the first time. Adoption does not always show up on the repository that gets the credit.
Neither of this week’s two biggest movers was on last week’s board at all, while five of these ten were. That collision is the reason to read a weekly ranking rather than a running one: the returning names tell you what is compounding, and the new ones tell you what changed.