Developer Adapters · Open Source
LedgerProof ships twenty-nine open-source adapter packages — covering every major LLM provider SDK, every major orchestration framework, every major cloud-managed AI service, and every major inference platform. Each adapter emits cryptographically signed Article 50 transparency receipts on the side-channel without modifying your response payload, without buffering streams, and without changing your latency profile. Local verification. No SaaS dependency. Foundation-stewarded protocol.
Frameworks & Orchestration
Orchestration frameworks
4 adaptersLangChain + LangGraph integration via a callback handler and a graph-node middleware. Receipts per LLM call and per graph node, including tool invocations.
LlamaIndex callback handler for RAG and chat. Receipts capture the retrieval step, the LLM step, and the final response — all on the side-channel.
Haystack pipeline component. Receipts emit per generator-node call in your pipeline, with retriever provenance preserved in the receipt context.
Microsoft Semantic Kernel kernel hook (Python). Receipts emit per InvokePromptAsync call. .NET port in active development.
Direct LLM Provider SDKs
LLM provider SDKs
12 adaptersOpenAI Python SDK client wrapper, @with_receipt decorator, and manual emission helper. Streaming and non-streaming paths both covered.
Anthropic Python SDK client wrapper, @with_receipt decorator, and tool-use support tuned for the Claude Agent SDK and multi-turn agent loops.
Mistral AI Python SDK client wrapper. Chat completions, function calling, and JSON-mode responses all emit receipts.
Codestral code-completion API wrapper. Receipts capture the completion request, the model identifier, and the response hash on the side-channel.
Cohere Python SDK client wrapper. Chat, generate, embed, and rerank endpoints all emit receipts tagged with the endpoint name.
AI21 Jamba and Jurassic API wrapper. Streaming completions supported; tool-use receipts include the function-call payload hash.
Aleph Alpha Pharia and Luminous SDK wrapper. Receipts tuned for sovereign-EU deployment patterns common in DACH financial services.
Google AI direct Gemini SDK wrapper (not Vertex). Supports the google-generativeai package; multimodal inputs are hashed deterministically.
xAI Grok SDK wrapper. Streaming completions and tool-use receipts; image-mode inputs hashed via canonical image-byte serialization.
DeepSeek API wrapper. Receipts emit for chat, reasoning, and code endpoints. Drop-in via the OpenAI-compatible interface.
Alibaba Qwen DashScope SDK wrapper. Receipts tuned for APAC-region deployments with PRC residency constraints.
Reka Core and Flash SDK wrapper. Multimodal inputs (audio, video, image) hashed via canonical serialization.
Cloud-Managed AI Services
Cloud-managed AI services
5 adaptersAWS Bedrock client wrapper covering Anthropic, Mistral, Cohere, AI21, Llama, and Amazon Titan model invocations. Streaming and tool-use both supported.
Google Vertex AI client wrapper. Receipts for Gemini, PaLM, and partner-hosted models. Stateful chat sessions hash conversation history canonically.
Azure OpenAI Service wrapper. Functions identical to openai-ledgerproof but tuned for the Azure deployment-name routing pattern and content-filter responses.
IBM watsonx.ai SDK wrapper. Granite and partner models supported. Receipts include the project ID and the model-deployment identifier for audit traceability.
Snowflake Cortex AI function wrapper. Receipts capture the warehouse identifier, the function name, and the row-level invocation pattern for table-bound AI calls.
Inference Platforms
Inference platforms
7 adaptersTogether AI inference API wrapper. Receipts emit per chat-completion and per embedding call. OpenAI-compatible endpoint also supported.
Groq LPU inference SDK wrapper. Low-latency anchoring path optimized for high-volume Groq deployments; receipts batch at the operator side without adding latency.
Replicate Python SDK wrapper. Receipts for image, audio, and video generation models include canonical hashes of binary outputs for Article 50(4) deepfake-disclosure use.
Hugging Face Inference API and Transformers pipeline wrapper. Receipts for hosted-endpoint calls; local-model invocations supported via the pipeline hook.
Fireworks AI inference SDK wrapper. OpenAI-compatible chat and completion endpoints; function calling and JSON-mode receipts both supported.
Cerebras Inference SDK wrapper. Very-high-throughput deployments use the batched-anchor configuration to keep Bitcoin OP_RETURN economics rational at scale.
Perplexity Sonar API wrapper. Receipts include the search-grounding citations as a deterministic hash so retrieved-source claims remain auditable.
Embeddings
Embedding providers
1 adapterVoyage AI embeddings SDK wrapper. Receipts capture the embedding-model identifier and the input-batch hash, useful when the embedding step itself is the AI-touched interaction.
Design Constraints
01 Side-channel emission
Adapters do not touch the bytes your application returns to the user. Receipts emit to a configurable sink — log, queue, webhook, or SIEM connector — on a separate path. Your existing integration tests, contract tests, and downstream consumers see nothing new in the response body.
02 Stream-aware signing
Streaming LLM responses are signed incrementally with a running SHA-256 over the token stream. The adapter does not buffer the response before forwarding it to your user. First-token latency is unchanged; receipt emission completes asynchronously after the stream closes.
03 Local verification
Every receipt is verifiable using only the Bitcoin chain, the published protocol public key, and the receipt itself. No call to LedgerProof servers required. The reference verifier runs in any browser. If LedgerProof Inc. disappears tomorrow, every receipt ever issued remains independently verifiable.
Start Here
If you're not sure which adapter fits your stack, start with langchain-ledgerproof — the LangChain callback handler works with any LLM provider your LangChain code already calls, and the integration is three lines of Python. Walk through it on the quickstart page.