Integrations

Every integration wires the same five things into the framework’s own extension points, always in the same design, so whoever learned one knows them all:

Primitive	What the adapter does
Context	Reads `context()` within the budget (150 ms on voice, 300 ms on the other views) and places the pack right after the agent’s instructions and the `turn_block` (other channels’ turns and the delta) at the end. The instructions stay first, so they remain the prefix the provider caches
Turns	Each customer utterance becomes `customer()`, each model answer becomes `agent()` with the usage the provider reported, and the end of the conversation becomes `end()`; the SDK’s queue delivers in the background
Tools	The three history tools (`search_customer_history`, `get_customer_timeline`, `open_history_item`) in the framework’s tool shape, with the customer bound outside the model’s reach
Verification	What the platform proves (the carrier’s STIR/SHAKEN attestation, a login, an OTP) goes to `verify()` before the first context
Handoff	When the framework transfers to another agent or to a person, `handoff()`, and the next agent opens the same conversation

None of them fails the agent: Niadra slow or down means no context, never an error. With agent memory on in the space, agent_memory=True (Python) or agentMemory: true (TypeScript) puts the agent’s own notes before the customer’s context and adds search_agent_memory (and remember) to the tools. In Python, each integration is an optional extra: pip install 'niadra[livekit]'. The base install keeps two dependencies (httpx and pydantic), and the framework is imported only inside the adapter. In TypeScript, each integration is a subpath of the package (import { ... } from "@niadra/sdk/livekit"), with the framework as an optional peer dependency. The TypeScript integrations ship with @niadra/sdk 0.3.0, ready on the repository’s main branch; it reaches npm when it is published, and until then the npm package is 0.1.1, without the subpaths.

The list

Integration	Python (extra)	TypeScript (subpath)	How it plugs in
LiveKit Agents	`livekit`	`livekit`	`NiadraAgent` and the `NiadraMemory` mixin: `llm_node` and `on_user_turn_completed`, session events
Pipecat	`pipecat`		`NiadraMemoryProcessor` between the user aggregator and the LLM
ElevenLabs Agents Platform	`elevenlabs`	`elevenlabs`	Initiation, server tool and post-call webhooks
Vapi	`vapi`	`vapi`	One server URL: `assistant-request`, `tool-calls`, transfers and `end-of-call-report`
Retell AI	`retell`	`retell`	Signed inbound webhook (dynamic variables), custom functions, call events; in TypeScript, the custom LLM websocket
WhatsApp Cloud API	`whatsapp`	`whatsapp`	Meta’s signed webhook into turns; the reply by the id Meta returns
Twilio	`twilio`	`twilio`	Voice with `StirVerstat` as proof, SMS and WhatsApp through Twilio
OpenAI Agents SDK	`openai-agents`	`openai-agents`	`call_model_input_filter`, `RunHooks` and function tools; in JavaScript, a `Session` and dynamic instructions
LangChain	`langchain`	`langchain`	A context runnable, a callback handler and `StructuredTool`s; LangGraph.js through `withNiadraContext()`
LangGraph	`langgraph`		`NiadraMiddleware` for `create_agent`, `pre_model_hook`
Vercel AI SDK		`ai-sdk`	A middleware for `wrapLanguageModel` and the kit as tools
Mastra		`mastra`	A processor for the model call and the kit as tools
Google ADK	`google-adk`	`google-adk`	`before_model` and `after_model` callbacks, `BaseTool`s; in TypeScript, `niadraAdk()`
Strands Agents	`strands`	`strands`	A `HookProvider` and tools; in TypeScript, `NiadraPlugin` (Node 22+)
Pydantic AI	`pydantic-ai`		A capability: dynamic instructions and a toolset
LlamaIndex	`llamaindex`	`llamaindex`	`NiadraMemory` and tools; in TypeScript, a `Memory`, a memory block and `FunctionTool`s
CrewAI	`crewai` (Python 3.11+)		Kickoff callbacks (`{niadra_context}`) and tools
Agno	`agno`		Dynamic instructions, `Function`s and run hooks
Microsoft Agent Framework	`agent-framework`		`NiadraContextProvider`
Semantic Kernel	`semantic-kernel`		`NiadraKernel`: agent thread, prompt filters and the `niadra` plugin
Haystack	`haystack` (3.2+)		`NiadraContext` and `NiadraReply` components, `NiadraAgentHooks` and `Tool`s
CAMEL-AI	`camel`		`NiadraMemory` (an `AgentMemory`) and `NiadraToolkit`
DSPy	`dspy`		`niadra_adapter()`, `NiadraModule` and `dspy.Tool`s
AG2	`ag2`		`NiadraAG2`: middleware and tools
Genkit		`genkit`	A model middleware for `generate({ use })` and Genkit tools
VoltAgent		`voltagent`	`onPrepareModelMessages`, `onEnd` and `onHandoff` hooks, and tools
Cloudflare Agents		`cloudflare-agents`	`niadraAgent()`: the AI SDK middleware with `ctx.waitUntil()`, tools, `prepare()` and `record()`
OpenAI and Azure OpenAI	`openai`	in the main package	`wrap()` of `chat.completions`
Anthropic	`anthropic`	`anthropic`	`wrap()` of `messages.create` and `messages.stream`
Amazon Bedrock	`bedrock`	`bedrock`	`wrap()` of `converse` and `converse_stream`
Google GenAI	`google-genai`	`google-genai`	`wrap()` of `generate_content` and its stream
LiteLLM	`litellm`		`completion()`, `acompletion()` and `NiadraLogger`
n8n		package `n8n-nodes-niadra`	A community node with six operations
Flowise		package `flowise-nodes-niadra`	A memory node and a tools node
Langflow	file in `integrations-extras/langflow`		Three components: context, reply and search
Dify	plugin in `integrations-extras/dify`		A tool provider: context, reply, history and agent memory

Not on the list: Zapier and Make (they require a developer account and a platform review; use the HTTP module with POST /v1/context) and observability tools (the OpenTelemetry ingestion already covers them). Any framework without an adapter uses the SDK directly: conversation.context(), customer(), agent() and tools().

MCP clients and coding assistants

The space’s MCP server connects to any client that speaks Streamable HTTP with custom headers: Claude Code, Claude Desktop, Cursor and VS Code. For the coding assistant to do the integration for you, there is the niadra skill.

Testing without a paid account

Every framework on the list is open source and installs without an account. Each adapter’s tests run in the SDKs’ CI against the framework’s real types, with the model, the STT and the TTS replaced by fakes, and Niadra on the niadra-mock emulator; the closed platforms (ElevenLabs, Vapi, Retell, Meta, Twilio) are tested with recorded payloads in the public format and signatures computed in the test itself. Every adapter has a runnable example in examples/ in the SDK repositories, the same code these pages show.

Coming from another memory

For those arriving from a vector-search memory, the name of each thing here; the guide Coming from Mem0 has the table call by call:

There	Here
`add()` on every message	`customer()` and `agent()` on the conversation, or `track()`; extraction runs at the end of the session, not on every message
`search()` before every answer	`context()`, pinned per conversation, and the search only when the pack does not answer
Feedback on a memory	`POST /v1/feedback`: retract or correct a fact, resolve an open item
Custom categories	Extraction predicates per space and per source
Direct import	`POST /v1/ingest/files`, CSV or JSONL
Export by schema	Continuous export in Parquet and `POST /v1/export`
Group conversation	`subjects` with a role, on an event with more than one subject
Async client	`AsyncNiadra`; the TypeScript SDK is asynchronous by nature
Multilingual	Extraction, context and search in any language; the pack labels in the space’s language; relative time in the search in Portuguese, English and Spanish
Expiration date	`valid_until` on the event and on the agent’s note
Procedural memory	Agent memory: explicit, reviewed, no personal data
`user_id`, `agent_id`, `run_id`	The customer’s handle (phone, e-mail, `wa_id`, app id), the source (the agent, with its vendor and key) and the conversation or task
`get_all()`, `history()` of a memory	The timeline (`timeline()`) and the `versions` of an opened item
Metadata filters on search	`filters`: period, `when`, channel, category, item kind, outcome and object
`delete()` and `delete_all()`	No delete tool within the model’s reach: erasure is `POST /v1/forget`, by a person or an `admin` key, with a receipt
`infer=False` (store without extracting)	A system event: structured data that enters without an AI model and changes its object’s state
Graph memory	Objects, links, accounts and partners: a typed graph, with no co-occurrence links
Hosted MCP client	The space’s MCP server, with the customer bound by the `subject_token`

Get started

Voice

Channels

Agent frameworks

Model SDKs

Automation platforms

MCP clients

The list

MCP clients and coding assistants

Testing without a paid account

Coming from another memory

Next steps

Quickstart

Agent memory

​The list

​MCP clients and coding assistants

​Testing without a paid account

​Coming from another memory

​Next steps

Quickstart

Agent memory

The list

MCP clients and coding assistants

Testing without a paid account

Coming from another memory

Next steps