[finding]
Providers sit at the center
Provider SDKs account for 23.1% of all normalized edges, and OpenAI SDK appears in 546 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-2026-03-29-methodology-v5 | | LENS: major, active, public OSS AI repos on GitHub | +--------------------------------------------------------------+
What major open source AI repos actually use.
This report reads directly from data/run-2026-03-29-methodology-v5 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
1,522
Candidate universe collected from GitHub discovery queries and manual seeds.
+ Final major AI repos
981
64.5% of discovered repos survive the full inclusion filter.
+ Normalized technology edges
5,187
979 repos map to at least one of 69 tracked technologies.
+ Median final repo stars
4,905
P75 is 12,557 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
303 of 979 technology-mapped repos (30.9%) use at least two tracked providers. The most common provider pairing is Anthropic SDK plus OpenAI SDK in 244 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
748 / 76.2%
User
233 / 23.8%
[language mix]
Python
540 / 55.0%
TypeScript
206 / 21.0%
Go
48 / 4.9%
Rust
42 / 4.3%
Jupyter Notebook
39 / 4.0%
+ coverage and method
Evidence tiers, validation audit, and where the map still undercounts
+ Direct-supported repos
938
Final repos with manifest, SBOM, import, or repo-identity evidence and no dependence on README-only fallback for coverage.
+ Fallback-supported repos
979
Final repos mapped once reviewed README fallback is allowed for otherwise unmapped repos.
+ README-only repos
41
Low-confidence fallback repos that remain explicit and separately inspectable in the publication artifact.
[1] Evidence tiers are now explicit
966 final repos have manifests and 837 have SBOM dependency evidence. 13 repos map only via canonical repo identity, 0 combine direct evidence with fallback signals, and 41 remain README-only.
[2] Normalization still leaves gaps
2 included repos (0.2%) 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 157 judge overrides were applied in this snapshot. The published set remains rule-first, but not purely rule-only. A seeded validation sample reviewed 99 final repos (10.1%) in addition to hardening. 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
925
readme only
41
repo identity only
13
unmapped
2
[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
2
Final repos that still have no normalized technology edge.
+ Unmapped dep gaps
2
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]
google
2,535
pydantic
2,523
@activepieces/
1,986
llama
1,975
nvidia
1,867
windows
1,708
[largest missing-edge repos]
FrigadeHQ/trench
unmapped dependency evidence • unmatched deps 114
1,620 stars
RunanywhereAI/RCLI
unmapped dependency evidence • unmatched deps 2
1,299 stars
[commodity and tooling backlog]
@rollup/
4,114
typescript
3,528
opentelemetry
3,366
github
3,337
python
3,318
serde
2,855
+ 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
18
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
93.9%
Repo included
93.9%
Identity mapped
83.7%
Third-party adoption
87.8%
Dependency evidence
95.9%
Negative excluded
100.0%
Holdout discovered
100.0%
[top benchmark gaps]
TGI
benchmark repo not discovered; no dependency evidence found for benchmark entity
score 140
Daytona
benchmark repo not discovered
score 100
DSPy
benchmark repo not discovered
score 100
llama.cpp
anchor entity is not covered by anchor seed inputs; repo included but canonical repo identity is not mapped
score 100
Ray Serve
anchor entity is not covered by anchor seed inputs; repo included but canonical repo identity is not mapped
score 100
+ robustness checks
How much the topline changes under stricter evidence and rule-only views
+ Rule-only final repos
972
Final-set size if the project uses raw rule outputs with no judge adjustment.
+ Judge-adjusted final repos
981
Published final-set size after hardening and validation overrides are applied.
+ Judge-changed finals
33
Repos whose final-set status differs between rule-only and published judge-adjusted views.
+ Direct-supported repos
938
Mapped repos using direct evidence or repo identity only.
+ Fallback lift
41
Additional mapped repos recovered only when reviewed README fallback is enabled.
+ Temporal delta
-4
Change in final repos relative to the baseline snapshot used for this validation pass.
Rule-only yields 972 final repos. Judge adjustment yields 981. Direct-only evidence maps 938 repos, and reviewed fallback lifts that to 979. Baseline comparison: /home/agent/oss-ai-stack-map/data/run-2026-03-25-repaired-v16
+ technology discovery
Post-filtered canonical candidates from the raw unmatched package graph
+ Raw candidate families
36
Unmatched technology families inferred from scraped dependency evidence before curation filters.
+ Graph nodes
36
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 the weight of the stack actually sits
+ Training and model ops
1,625
[31.3% of edges]
Transformers
377 / 38.4%
PyTorch
324 / 33.0%
Hugging Face Hub
277 / 28.2%
Tokenizers
245 / 25.0%
Accelerate
208 / 21.2%
+ Providers and access
1,196
[23.1% of edges]
OpenAI SDK
563 / 57.4%
Anthropic SDK
290 / 29.6%
Google GenAI SDK
215 / 21.9%
LiteLLM
128 / 13.0%
+ Orchestration and agents
857
[16.5% of edges]
LangChain
246 / 25.1%
LangChain OpenAI Integration
146 / 14.9%
LangGraph
115 / 11.7%
LangChain Anthropic Integration
58 / 5.9%
LangChain Google GenAI Integration
55 / 5.6%
+ Ai Developer Tools And Sdk Families
488
[9.4% of edges]
Model Context Protocol
384 / 39.1%
Vercel AI SDK
102 / 10.4%
code2prompt
1 / 0.1%
Grafbase
1 / 0.1%
+ Retrieval and vector storage
368
[7.1% of edges]
Chroma
93 / 9.5%
Qdrant
88 / 9.0%
pgvector
52 / 5.3%
Milvus
46 / 4.7%
LanceDB
43 / 4.4%
+ Serving and local runtimes
265
[5.1% of edges]
Ollama
129 / 13.1%
vLLM
59 / 6.0%
Ray Serve
47 / 4.8%
llama.cpp
18 / 1.8%
SGLang
10 / 1.0%
+ UI and app frameworks
212
[4.1% of edges]
Gradio
119 / 12.1%
Streamlit
81 / 8.3%
Chainlit
12 / 1.2%
+ Observability
70
[1.3% of edges]
Langfuse
42 / 4.3%
Logfire
17 / 1.7%
Arize Phoenix
5 / 0.5%
Weave
4 / 0.4%
Evidently
1 / 0.1%
+ Sandbox And Isolated Execution
36
[0.7% of edges]
E2B
27 / 2.8%
Daytona
9 / 0.9%
+ Browser And Computer Use Infra
35
[0.7% of edges]
Browserbase
22 / 2.2%
Browser Use
12 / 1.2%
Steel Browser
1 / 0.1%
+ Evaluation and guardrails
32
[0.6% of edges]
Ragas
15 / 1.5%
Guardrails
6 / 0.6%
DeepEval
5 / 0.5%
Promptfoo
4 / 0.4%
NeMo Guardrails
2 / 0.2%
+ Runtime And Agent Deployment
3
[0.1% of edges]
Cloudflare Agents
3 / 0.3%
+ top technologies
The most repeated building blocks
OpenAI SDK
563 / 57.4%
Model Context Protocol
384 / 39.1%
Transformers
377 / 38.4%
PyTorch
324 / 33.0%
Anthropic SDK
290 / 29.6%
Hugging Face Hub
277 / 28.2%
LangChain
246 / 25.1%
Tokenizers
245 / 25.0%
Google GenAI SDK
215 / 21.9%
Accelerate
208 / 21.2%
+ provider and segment mix
Who dominates, and what gets built
[provider prevalence]
OpenAI SDK
546 / 55.7%
Anthropic SDK
269 / 27.4%
Google GenAI SDK
246 / 25.1%
[primary segments]
Serving runtime
297 / 30.3%
Training and finetuning
283 / 28.8%
Orchestration framework
122 / 12.4%
Vector and retrieval infra
99 / 10.1%
Agent application
73 / 7.4%
RAG and search app
41 / 4.2%
+ repeated combinations
The strongest co-usage patterns
PyTorch :: Transformers
Shared by major OSS AI repos in the normalized graph.
[269]
Anthropic SDK :: OpenAI SDK
Shared by major OSS AI repos in the normalized graph.
[264]
Model Context Protocol :: OpenAI SDK
Shared by major OSS AI repos in the normalized graph.
[243]
OpenAI SDK :: Transformers
Shared by major OSS AI repos in the normalized graph.
[226]
Hugging Face Hub :: OpenAI SDK
Shared by major OSS AI repos in the normalized graph.
[213]
Hugging Face Hub :: Tokenizers
Shared by major OSS AI repos in the normalized graph.
[205]
+ at a glance
A few stable signals from the snapshot
+ Serious repos
1,219
80.1%
+ AI-relevant repos
1,084
71.2%
+ Repos with normalized techs
979
99.8%
+ 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 7 communities with modularity 0.1072. 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 68.7%
46 technologies
top categories
Orchestration (17)Retrieval (7)Providers (4)
top technologies
Anthropic SDKArize PhoenixAutoGenBrowser UseBrowserbaseChainlit
exemplar repos
ComposioHQ/composiolangflow-ai/langflowag-ui-protocol/ag-uiArize-ai/phoenixtraceloop/openllmetry
[community 2]
Training And Inference Core
Model-training and inference-runtime technologies clustered around finetuning, serving, and heavyweight model execution.
share 23.9%
16 technologies
top categories
Training (8)Serving (5)UI (1)
top technologies
AccelerateBentoMLDeepSpeedGradioHeliconeHugging Face Hub
exemplar repos
run-llama/llama_indexmicrosoft/LMOpsray-project/rayOpenBMB/MiniCPMpostgresml/postgresml
[community 3]
Mixed Stack Cluster
A cross-linked family of technologies without a single dominant role.
share 1.5%
1 technologies
top categories
ai_developer_tools_and_sdk_families (1)
top technologies
Grafbase
exemplar repos
grafbase/grafbase