Built by Cognipeer, in production, in the open

An open AI platform,
end to end.

Two flagships do the heavy lifting — Console as the multi-tenant control plane, Agent SDK as the deterministic agent runtime. A growing toolkit of libraries plugs into both. We use it ourselves. You can ship on the same stack.

Start with ConsoleBuild with Agent SDK
  Self-host or hostedAny LLM providerTypeScript-firstAGPL & permissive licenses
cognipeer.stack  ·  one platform, two flagships
// The shape of your AI stack on Cognipeer
 
interface CognipeerStack {
// flagship: the control plane
console: {
gateway: "openai-compatible",
providers: "openai | anthropic | azure | bedrock | vertex | ...",
vectors: "any", rag: true, tracing: true,
guardrails: true, prompts: true, tenancy: "multi",
},
// flagship: the runtime
agentSdk: {
loop: "message-first", planning: true,
reasoning: "fast | balanced | deep | research",
multiAgent: true, tracing: "otel",
},
// supporting libraries plug into both
extras: "console-sdk · client-sdk · agent-server · chat-ui · ...",
}
BUILT BY US · USED BY US · OPEN FOR YOU

Not a side project. The same stack runs Cognipeer.

Every library here exists because we needed it ourselves. We open-source the parts other teams can reuse — and keep shipping new ones as our own workloads grow.

In production at
Cognipeer
our hosted & enterprise deployments
Customer agents shipped on it
hundreds
across banking, telco, retail, energy
Inference routed monthly
millions
through Console's LLM gateway
Open-source toolkit
growing · steadily
new libraries as our own workloads grow
Powering production AI for teams in
banking·telco·insurance·retail·airlines·healthcare·energy·public sector
Flagships

Two products carry the platform.

Console is everything you need to run AI in production. Agent SDK is everything you need to build agents that survive contact with production. Use one, both, or layer the rest of the toolkit on top.

Flagship · Control plane

Cognipeer Console

@cognipeer/console  ·  AGPL-3.0  ·  self-host or hosted

An open, multi-tenant AI gateway. Operate inference, vector stores, RAG, tracing, guardrails, prompts, and quotas behind one console — with tenant isolation built in.

LLM GatewayOpenAI-compatible chat & embeddings across OpenAI, Anthropic, Bedrock, Vertex, Azure, vLLM, Ollama, and more.
Vector store & RAGMulti-provider vector operations, document ingestion, chunking, embedding, and retrieval.
Tracing & tracing sinksBatch and streaming ingest with thread correlation and OTel-compatible IDs.
Guardrails & promptsPII detection, content moderation, prompt shields, versioned templates with environments.
Quotas & isolationMulti-dimensional quotas, rate limits, and per-tenant databases out of the box.
SQLite-by-defaultDocker / HelmMulti-tenant
Flagship · Runtime

Agent SDK

@cognipeer/agent-sdk  ·  npm  ·  Node 18.17+

Lightweight, message-first agent runtime. Tool calls stay transparent, long histories summarize without losing evidence, and one config controls planning, reasoning, and reflection.

Native provider layerCall OpenAI, Anthropic, Azure, Bedrock, Vertex, or any OpenAI-compatible API directly. No LangChain required.
Reasoning + reflectionOne reasoning config switches provider-native reasoning and post-tool reflections.
Runtime profilesFast, balanced, deep, or research presets — real tradeoff bundles for context, depth, memory.
Plans, summaries, recoveryDurable plan on state.plan; archived tool outputs still retrievable via get_tool_response.
Multi-agent & structured outputWrap agents as tools, hand off mid-run, return Zod-validated JSON deterministically.
Message-firstZero-graphOTel tracing
How they fit, end to end
Your app
Web
Mobile
Backend
Internal tools
Build with
Agent SDK
Chat UI
Client SDK
Console SDK
Agent Server
Run on
Cognipeer Console
LLM gateway · vector stores · RAG · prompt mgmt · guardrails · tracing · quotas · multi-tenancy
Talks to
OpenAI
Anthropic
Azure
Bedrock
Vertex
vLLM
Ollama
any OpenAI-compatible
Supporting toolkit

The libraries that plug into both.

Targeted libraries that extend Console and Agent SDK at specific points in the stack — and a growing set of utilities we open-source as we need them.

Console SDK

TypeScript SDK for Cognipeer Console
Official TypeScript/JavaScript SDK for Cognipeer Console. Provides type-safe access to chat completions, embeddings, vector operations, file management, agent tracing, and guardrails.
@cognipeer/console-sdk

Client SDK

Studio Integration SDK
Official JavaScript and TypeScript SDK for Cognipeer Studio. Connect applications to Studio conversations, channels, peers, users, and client-side tool execution from the browser or server.
@cognipeer/sdk

Agent Server

Agent Serving Infrastructure
A production-ready serving layer that registers, versions, and exposes multiple AI agents as secure, streaming-enabled REST services with Swagger/OpenAPI documentation.
@cognipeer/agent-server

Chat UI

AI Conversational Interface Kit
A React component library for building modern, real-time conversational interfaces powered by AI agents. ChatGPT-style experience with streaming, tool transparency, and full customisation.
@cognipeer/chat-ui

to-markdown

File → Markdown converter
A versatile, TypeScript-first utility for converting PDF, DOCX, HTML, Excel, CSV, and more into clean Markdown — ready for RAG pipelines and LLM context windows.
@cognipeer/to-markdown

graph-builder

Reasoning graphs & document flows
An API-first TypeScript library that turns text sources into a queryable graph. Build it from texts, providers, or change streams — and persist, update, or replay the result.
@cognipeer/graph-builder

Actively expanding. New packages ship when our own work needs them — adapters, components, evals, recipes. Watch the GitHub org for the next release.

Browse on GitHub
Quickstart

Pick the flagship that matches your day.

Need to run AI for an org? Start with Console. Building an agent against any LLM? Start with the SDK.

PATH A · Platform team

Deploy Cognipeer Console

Self-host the open AI gateway. SQLite-by-default — no external database required.

runs anywhere with Node 20+
$  git clone https://github.com/Cognipeer/console
$  cd console && npm install
$  cp .env.example .env.local
$  npm run dev
or with Docker
$docker compose up -d
Read the Console guide
PATH B · App developer

Build with Agent SDK

One install, one config, any LLM. Planning, reflection, and tracing are built in.

$npm install @cognipeer/agent-sdk zod
1import {
2  createSmartAgent, createTool,
3  createProvider, fromNativeProvider,
4} from "@cognipeer/agent-sdk";
5 
6const model = fromNativeProvider(
7  createProvider({ provider: "openai", apiKey: process.env.OPENAI_API_KEY! }),
8  { model: "gpt-4o" },
9);
10 
11const agent = createSmartAgent({
12  model,
13  tools: [searchKb, openTicket],
14  runtimeProfile: "balanced",
15  tracing: { enabled: true },
16});
17 
18const result = await agent.invoke({
19  messages: [{ role: "user", content: "Where is order #4187?" }],
20});
Read the SDK guide
Recipes

Patterns we've shipped — now in code.

Production-tested recipes for agents, RAG, governance, and embedded chat. Each one shows the package mix and a starting point you can copy.

All recipes
Recipe · 6 min read

HR Self-Service Widget

Embed a Studio-powered HR assistant into an employee portal for policy Q&A, leave requests, onboarding help, and document routing.
Client SDKChat · Widget
Recipe · 6 min read

Expense Upload & Approval Widget

Let employees upload receipts, extract fields, and submit expense requests through a Studio-powered finance widget embedded in your product.
Client SDKChat · Widget
Recipe · 8 min read

Mobile Peer Embedded In The App

Run a Studio peer directly inside a mobile app so it can answer users and trigger native app actions such as screen navigation, camera capture, and task execution.
Client SDKChat · Mobile
Recipe · 7 min read

Website Product Advisor Widget

Embed a Studio peer into a public website to qualify visitors, answer pricing questions, and trigger product actions without building a separate bot backend.
Client SDKChat · Widget
Recipe · 8 min read

Quota-Aware LLM Gateway

Run application traffic through Console with project-scoped quotas, model routing, and request-level visibility for cost control.
Console + Console SDKPlatform · Gateway
Recipe · 9 min read

Vector RAG Operations Control Plane

Operate RAG pipelines through Console by combining file ingestion, vector index management, embeddings, and chat retrieval in one control surface.
Console + Console SDKRAG · Platform
Roadmap

Public, opinionated, regularly updated.

What we just shipped, what's next, and what we're researching. Vote, comment, or open an RFC.

Shipped

Past 30 days

  • agent-sdk: runtime profiles
    fast / balanced / deep / research presets — real tradeoff bundles.
  • console: per-project quotas
    Multi-dimensional limits with automatic provider failover.
  • chat-ui: tool-call inspector
    Inline panel for inputs/outputs, latency, partial sink failures.
In progress

This quarter

  • Evals harness
    Offline grading and scenario replay over Console traces.
  • MCP tool gateway
    Quota-aware adapter for MCP servers, exposed through Console.
  • Edge runtime for Agent SDK
    Workers-compatible build for low-latency deployments.
Exploring

Next up

  • Visual agent planner
    DAG-style preview of agent reasoning, served from Console.
  • Policy DSL
    Declarative guardrails composable across tenants and agents.
  • Mobile Chat UI parity
    React Native build of the conversational kit.
Open the public board
Changelog

Released this week

Granular notes for every library, plus migration tips when an API shifts.

All releases
May 18 · 2 days ago
agent-sdk

Reasoning presets and durable plans

New reasoning config switches between fast, balanced, deep, and research presets — real tradeoff bundles for context budget, delegation depth, and memory policy. Plans now live on state.plan instead of transient UI events.

May 14 · 6 days ago
console

Per-project quotas + provider routing v2

Define burst & sustained limits per project, with automatic routing across providers when a primary degrades. The gateway exposes per-tenant headers and surfaces overage events to webhooks.

May 09 · 11 days ago
chat-ui

Tool-call inspector + attachment streaming

A first-class panel that shows tool input/output, latency, and partial sink failures inline. Attachments stream incrementally so users see images and files appear as they're generated.

Open infrastructure for serious AI

Build on the same stack we use to ship Cognipeer.

Self-host Console, install Agent SDK, layer the rest of the toolkit on top. Or skip the operations and use Cognipeer hosted — same APIs, same SDKs.