[finding]
Providers sit at the center
Provider SDKs account for 22.9% of all normalized edges, and OpenAI SDK appears in 549 final repos. The provider layer is the most common and the most connective part of the stack.
+--------------------------------------------------------------+ | OSS AI STACK MAP :: SNAPSHOT REPORT | | SOURCE: data/run-2026-03-31-publication-v8 | | LENS: major, active, public OSS AI repos on GitHub | +--------------------------------------------------------------+
What major open source AI repos actually use.
This report reads directly from data/run-2026-03-31-publication-v8 and summarizes the current stack choices across the project’s final GitHub AI set.
Study frame: GitHub-only, public, non-fork, non-archived, active within 1 month, and at least 1,000 stars. Published stack edges come from manifests, SBOMs, bounded import fallback, repo identity, and reviewed README fallback when an included repo would otherwise remain unmapped.
+ Discovered repos
1,525
Candidate universe collected from GitHub discovery queries and manual seeds.
+ Final major AI repos
983
64.5% of discovered repos survive the full inclusion filter.
+ Normalized technology edges
5,255
983 repos map to at least one of 77 tracked technologies.
+ Median final repo stars
4,911
P75 is 12,580 stars across the published final set.
+ main findings
The modern OSS AI stack is provider-first, orchestration-heavy, and surprisingly multi-vendor.
[finding]
Multi-provider stacks are common
306 of 983 technology-mapped repos (31.1%) use at least two tracked providers. The most common provider pairing is Anthropic SDK plus OpenAI SDK in 246 repos. That is followed by Google GenAI SDK plus OpenAI SDK in 218 repos. Major OSS projects are not clustering around a single vendor.
[finding]
The ecosystem is backend-heavy
Training, orchestration, providers, and retrieval dominate the graph. Evaluation and observability remain comparatively thin, with only 32 guardrail/eval edges and 70 observability edges.
+ stack profile
A compact read on the default stack shape
Inference from the aggregate counts: the modal major OSS AI repo in this snapshot is organization-owned, Python-first, anchored on a provider SDK, often layers in Hugging Face training tools, and then adds orchestration, retrieval, and a lightweight UI shell.
[who ships it]
Organization
750 / 76.3%
User
233 / 23.7%
[language mix]
Python
542 / 55.1%
TypeScript
206 / 21.0%
Go
48 / 4.9%
Rust
42 / 4.3%
Jupyter Notebook
39 / 4.0%
+ entity mapping
Who appears to steward repos, and which companies sit behind tracked technologies
This layer is separate from stack normalization. Repo steward mapping currently uses curated exact repo-name and GitHub org matches only. Technology vendor mapping is curated in config and should be read as product stewardship, not proof that every adopting repo is company-backed.
+ Tracked entities
24
Curated company, startup, foundation, and individual entities loaded from the entity registry.
+ Steward-mapped repos
54
5.5% of final repos map to a steward entity via repo-name or GitHub org evidence.
+ Vendor-mapped repos
856
87.1% of final repos use at least one technology tied to a curated vendor entity across 34 tracked technologies.
[top repo stewards]
Hugging Face
company
10 / 1.0%
LangChain
startup
8 / 0.8%
Google
company
7 / 0.7%
OpenAI
company
5 / 0.5%
Meta
company
4 / 0.4%
Qdrant
startup
3 / 0.3%
[top technology vendors]
OpenAI
2 mapped technologies
566 / 57.6%
Hugging Face
5 mapped technologies
487 / 49.5%
Anthropic
1 mapped technologies
292 / 29.7%
LangChain
5 mapped technologies
256 / 26.0%
Google
1 mapped technologies
215 / 21.9%
Ollama
1 mapped technologies
129 / 13.1%
[entity types]
startup
16
company
8
[repo steward confidence]
high
54
+ coverage and method
Evidence tiers, validation audit, and where the map still undercounts
+ Direct-supported repos
941
Final repos with manifest, SBOM, import, or repo-identity evidence and no dependence on README-only fallback for coverage.
+ Fallback-supported repos
983
Final repos mapped once reviewed README fallback is allowed for otherwise unmapped repos.
+ README-only repos
42
Low-confidence fallback repos that remain explicit and separately inspectable in the publication artifact.
[1] Evidence tiers are now explicit
968 final repos have manifests and 839 have SBOM dependency evidence. 13 repos map only via canonical repo identity, 0 combine direct evidence with fallback signals, and 42 remain README-only.
[2] Normalization still leaves gaps
0 included repos (0.0%) have no normalized technology edge, so graph-like analysis describes the mapped subset, not the entire final population.
[3] Validation is now an audit layer
651 repos were judge-reviewed and 158 judge overrides were applied in this snapshot. The published set remains rule-first, but not purely rule-only. The validation audit reviewed 119 repos, changed 17 decisions, excluded 7 repos from the sampled final set, and estimates a false-positive rate of 5.9% with a 95% interval of 2.9% to 11.7%.
[evidence profiles]
direct only
928
readme only
42
repo identity only
13
[validation sample by segment]
Serving runtime
42
Training and finetuning
25
Vector and retrieval infra
15
Orchestration framework
13
Agent application
10
RAG and search app
5
+ research gaps
Where normalization still misses stack evidence
+ Missing-edge repos
0
Final repos that still have no normalized technology edge.
+ Unmapped dep gaps
0
Missing-edge finals that do have dependency evidence, but none of it normalizes yet.
+ No-dependency gaps
0
Missing-edge finals that have no dependency evidence at all.
+ AI-specific prefixes
10
High-signal unresolved package families that still look AI-stack specific.
+ Commodity prefixes
10
Generic tooling and ecosystem noise separated from the research backlog.
+ Vendor-like repos
10
Repos that look vendor-related but are not mapped to a canonical identity.
[ai-specific unmatched prefixes]
llama
1,975
nvidia
1,867
langchain
880
datasets
608
@crawlee/
463
@langchain/
385
[largest missing-edge repos]
No missing-edge repos in this snapshot.
[commodity and tooling backlog]
typescript
3,547
opentelemetry
3,418
github
3,407
python
3,340
serde
2,863
@coze-arch/
2,612
+ benchmark recall
How well the map covers the current benchmark panel, holdout set, and negative controls
+ Positive entities
49
Tracked OSS AI stack entities used for recall measurement across this snapshot.
+ Negative controls
10
Discovered candidate repos that should stay out of the final set and act as a precision guardrail.
+ Holdout entities
12
Positive entities reserved as a holdout slice instead of the main tuning panel.
+ Failed thresholds
0
Recall metrics currently below configured minimums.
+ Prioritized gaps
15
Benchmarks needing the next registry or discovery fixes.
+ Total benchmark entries
59
Combined positives and negatives now tracked by the benchmark report.
[coverage rates]
Repo discovered
100.0%
Repo included
100.0%
Identity mapped
98.0%
Third-party adoption
87.8%
Dependency evidence
98.0%
Negative excluded
100.0%
Holdout discovered
100.0%
[top benchmark gaps]
LanceDB
repo included but canonical repo identity is not mapped
score 80
DingoDB
no dependency evidence found for benchmark entity
score 40
Daytona
benchmark repo discovered only via exact seed
score 30
DSPy
benchmark repo discovered only via exact seed
score 30
llama.cpp
benchmark repo discovered only via exact seed
score 30
+ robustness checks
How much the topline changes under stricter evidence and rule-only views
+ Rule-only final repos
975
Final-set size if the project uses raw rule outputs with no judge adjustment.
+ Judge-adjusted final repos
983
Published final-set size after hardening and validation overrides are applied.
+ Judge-changed finals
34
Repos whose final-set status differs between rule-only and published judge-adjusted views.
+ Direct-supported repos
941
Mapped repos using direct evidence or repo identity only.
+ Fallback lift
42
Additional mapped repos recovered only when reviewed README fallback is enabled.
+ Temporal delta
0
Change in final repos relative to the baseline snapshot used for this validation pass.
Rule-only yields 975 final repos. Judge adjustment yields 983. Direct-only evidence maps 941 repos, and reviewed fallback lifts that to 983. Baseline comparison: /home/agent/oss-ai-stack-map/data/run-2026-03-31-publication-v7
+ technology discovery
Post-filtered canonical candidates from the raw unmatched package graph
+ Raw candidate families
35
Unmatched technology families inferred from scraped dependency evidence before curation filters.
+ Graph nodes
35
Package-family nodes in the projected co-usage graph.
+ Filtered candidates
0
Canonical vendor, product, or package-family candidates that survive the registry filter.
No post-filtered canonical discovery candidates available for this snapshot.
+ registry suggestions
Canonical vendor, product, and package-family suggestions
+ Suggestions
0
Candidates produced after filtering already-covered families and suppressing abstract capability labels.
+ LLM reviewed
0
Optional OpenAI registry reviews attached to candidate suggestions.
+ Shown here
0
Top suggestion rows displayed in this report.
No registry suggestions available for this snapshot.
+ modern ai stack layers
Where normalized stack usage concentrates
These cards rank categories by normalized repo-tech edges, not by architectural importance. Each card now shows both edge share and repo prevalence across the 983 final repos. Very thin categories are summarized separately below.
+ Model frameworks and HF stack
1,635
normalized repo-tech edges
[31.1% of edges]
[523 repos / 53.2%]
Transformers
378 / 38.5%
PyTorch
325 / 33.1%
Hugging Face Hub
279 / 28.4%
Tokenizers
247 / 25.1%
Accelerate
209 / 21.3%
+ Providers and access
1,203
normalized repo-tech edges
[22.9% of edges]
[621 repos / 63.2%]
OpenAI SDK
566 / 57.6%
Anthropic SDK
292 / 29.7%
Google GenAI SDK
215 / 21.9%
LiteLLM
130 / 13.2%
+ Orchestration and agents
861
normalized repo-tech edges
[16.4% of edges]
[304 repos / 30.9%]
LangChain ecosystem
LangChain 247, LangChain OpenAI Integration 146, LangChain Anthropic Integration 58, LangChain Google GenAI Integration 55, LangGraph 115
256 / 26.0%
LlamaIndex
51 / 5.2%
OpenAI Agents
34 / 3.5%
CrewAI
32 / 3.3%
Instructor
30 / 3.1%
+ Protocols and developer SDKs
491
normalized repo-tech edges
[9.3% of edges]
[429 repos / 43.6%]
Model Context Protocol
386 / 39.3%
Vercel AI SDK
103 / 10.5%
code2prompt
1 / 0.1%
Grafbase
1 / 0.1%
+ Retrieval and vector storage
369
normalized repo-tech edges
[7.0% of edges]
[191 repos / 19.4%]
Chroma
93 / 9.5%
Qdrant
88 / 9.0%
pgvector
52 / 5.3%
Milvus
46 / 4.7%
LanceDB
43 / 4.4%
+ Serving and local runtimes
269
normalized repo-tech edges
[5.1% of edges]
[214 repos / 21.8%]
Ollama
129 / 13.1%
vLLM
59 / 6.0%
Ray Serve
47 / 4.8%
llama.cpp
21 / 2.1%
SGLang
10 / 1.0%
+ UI and app frameworks
212
normalized repo-tech edges
[4.0% of edges]
[183 repos / 18.6%]
Gradio
119 / 12.1%
Streamlit
81 / 8.2%
Chainlit
12 / 1.2%
+ Sandbox and isolated execution
73
normalized repo-tech edges
[1.4% of edges]
[56 repos / 5.7%]
E2B
34 / 3.5%
Modal
19 / 1.9%
Daytona
13 / 1.3%
Runloop
4 / 0.4%
Vercel Sandbox
2 / 0.2%
+ Observability
70
normalized repo-tech edges
[1.3% of edges]
[64 repos / 6.5%]
Langfuse
42 / 4.3%
Logfire
17 / 1.7%
Arize Phoenix
5 / 0.5%
Weave
4 / 0.4%
Evidently
1 / 0.1%
+ Browser and computer use infra
37
normalized repo-tech edges
[0.7% of edges]
[34 repos / 3.5%]
Browserbase
22 / 2.2%
Browser Use
12 / 1.2%
Hyperbrowser
2 / 0.2%
Steel Browser
1 / 0.1%
+ Evaluation and guardrails
32
normalized repo-tech edges
[0.6% of edges]
[30 repos / 3.1%]
Ragas
15 / 1.5%
Guardrails
6 / 0.6%
DeepEval
5 / 0.5%
Promptfoo
4 / 0.4%
NeMo Guardrails
2 / 0.2%
+ thinly tracked layers
Visible, but not broad enough for a primary card
Runtime and agent deployment
3 normalized repo-tech edges across 3 repos. Thin categories stay listed here until coverage is broad enough for a full card.
Cloudflare Agents 3
[0.1% edges • 0.3% repos]
+ top technologies
The most repeated building blocks
OpenAI SDK
566 / 57.6%
Model Context Protocol
386 / 39.3%
Transformers
378 / 38.5%
PyTorch
325 / 33.1%
Anthropic SDK
292 / 29.7%
Hugging Face Hub
279 / 28.4%
LangChain ecosystem
LangChain 247, LangChain OpenAI Integration 146, LangChain Anthropic Integration 58, LangChain Google GenAI Integration 55, LangGraph 115
256 / 26.0%
Tokenizers
247 / 25.1%
Google GenAI SDK
215 / 21.9%
Accelerate
209 / 21.3%
+ provider and segment mix
Who dominates, and what gets built
[provider prevalence]
OpenAI SDK
549 / 55.8%
Anthropic SDK
271 / 27.6%
Google GenAI SDK
246 / 25.0%
[primary segments]
Serving runtime
297 / 30.2%
Training and finetuning
284 / 28.9%
Orchestration framework
124 / 12.6%
Vector and retrieval infra
99 / 10.1%
Agent application
73 / 7.4%
RAG and search app
40 / 4.1%
+ repeated combinations
The strongest co-usage patterns
PyTorch :: Transformers
Shared by major OSS AI repos in the normalized graph.
[270]
Anthropic SDK :: OpenAI SDK
Shared by major OSS AI repos in the normalized graph.
[266]
Model Context Protocol :: OpenAI SDK
Shared by major OSS AI repos in the normalized graph.
[245]
OpenAI SDK :: Transformers
Shared by major OSS AI repos in the normalized graph.
[227]
Hugging Face Hub :: OpenAI SDK
Shared by major OSS AI repos in the normalized graph.
[215]
Hugging Face Hub :: Tokenizers
Shared by major OSS AI repos in the normalized graph.
[207]
+ at a glance
A few stable signals from the snapshot
+ Serious repos
1,222
80.1%
+ AI-relevant repos
1,086
71.2%
+ Repos with normalized techs
983
100.0%
+ Median mapped tech count
4
Across the technology-connected subset only.
+ graph structure
Which technologies sit at the center of the mapped stack
These visuals summarize the technology-connected subset of the final population. Eigenvector highlights the core hubs, betweenness isolates bridge technologies, repo degree shows stack breadth per mapped repo, and category mixing shows which layers of the stack actually co-occur.
[top eigenvector technologies]
High-eigenvector technologies are not just common. They sit next to other highly connected technologies and define the graph’s center of gravity.
[betweenness vs prevalence]
Points higher on the chart bridge otherwise different tool combinations. Point size reflects weighted co-occurrence strength.
[repo degree distribution]
This is the distribution of tracked technologies per mapped repo, not the full final set including no-edge repos.
[category mixing heatmap]
Weighted co-occurrence between the biggest stack categories. Darker cells indicate heavier cross-category coupling in the technology projection.
+ community structure
How the technology graph breaks into stack families
Greedy modularity on the technology projection found 8 communities with modularity 0.1071. Lower modularity means the graph is still heavily cross-linked, so these are useful stack families rather than cleanly isolated islands.
[community 1]
Retrieval And App Surface
Vector storage, retrieval plumbing, and lightweight app frameworks that often sit at the presentation edge of AI systems.
share 67.1%
49 technologies
top categories
Orchestration (17)Retrieval (7)Providers (4)
top technologies
Anthropic SDKArize PhoenixAutoGenBrowser UseBrowserbaseChainlit
exemplar repos
ComposioHQ/composiolangflow-ai/langflowrun-llama/llama_indexag-ui-protocol/ag-uiArize-ai/phoenix
[community 2]
Training And Inference Core
Model-training and inference-runtime technologies clustered around finetuning, serving, and heavyweight model execution.
share 23.3%
17 technologies
top categories
Training (8)Serving (6)sandbox_and_isolated_execution (1)
top technologies
AccelerateBentoMLDeepSpeedGradioHugging Face Hubllama.cpp
exemplar repos
run-llama/llama_indexmicrosoft/LMOpsray-project/raypydantic/pydantic-aitensorzero/tensorzero
[community 3]
Mixed Stack Cluster
A cross-linked family of technologies without a single dominant role.
share 2.7%
2 technologies
top categories
sandbox_and_isolated_execution (1)Observability (1)
top technologies
Cloudflare ContainersHelicone
exemplar repos
Helicone/helicone