KDCube is a self‑hosted multi‑tenant, multi‑user platform + SDK for building custom AI assistants, copilots, and agentic apps.
Run a copilot for your customers, not just for yourself — use the built‑in ReAct workflow or bring your own.
KDCube provides the runtime, streaming protocol, tools, memory, and operations stack so you can ship AI products with production‑grade reliability.
Highlights
- Full stack: from streaming protocols to tool execution, memory, economics, and ops.
- Agent‑first: a ready‑made, versatile ReAct‑style agent that can be extended or replaced.
- No tool‑calling lock‑in: the agent runs on plain prompt/completion while fully emulating tool‑calling (and more).
- Built for real apps: multi‑tenant isolation, backpressure, rate limits, and observability.
- Channeled streaming + live widgets: animate UX with streaming channels and custom widgets.
- Provenance by default: source pools and citations to prove how answers were built.
- Feedback‑aware: user feedback is captured and can be fed back into workflows.
Install the bootstrap CLI and launch the setup wizard:
pipx install kdcube-cli
kdcube-setupAlternative (pip):
pip install -kdcube-cli
kdcube-setupPrerequisites: Python 3.9+, Git, and Docker (Docker Desktop on macOS/Windows).
Runtime & streaming
- Streaming chat over SSE / REST / Socket.IO with step/delta/status events
- Fine‑grained streaming channels (answer, reasoning, artifacts, subsystem payloads)
- Session‑aware relay + fan‑out for multiple tabs/clients
Agent capabilities
- Versatile ReAct‑style solver (planning, tool‑first/code‑first flows)
- Skills + tools (local + MCP) with easy custom wiring
- Built‑in web search and citations pipeline
Execution & artifacts
- Isolated code execution (Docker + Fargate)
- Attachments + generated artifacts with storage + indexing
- Antivirus scanning for uploads
- Dynamic widgets (interactive timeline banners + live content: web search/fetch, exec panels, bundle‑driven panels)
Memory & context
- Turn/context memories, conversation memories, retrieval
- Source pools per conversation (Perplexity‑style traceability)
- Structured sources + citations (with in‑stream rendering)
Operations & safety
- Multi‑tenant / multi‑project isolation
- Gateway: auth, rate limits, backpressure, circuit breakers
- Economics & accounting (usage, budgets, rate limits)
- Monitoring + metrics service for autoscaling
- Feedback system (user feedback signals integrated into workflow)
- Role‑based event filtering (stream only what each role is allowed to see)
- Dynamic bundle UIs (React interfaces exposed by bundles, authorization‑guarded)
- Agent runtime: ReAct v2 + planning and tool/code orchestration
- Streaming protocol: step/delta/status + widget channels
- Tools & skills: MCP + custom tools
- Context & memory: turn memories, signals, retrieval
- Execution runtime: isolated code execution
- Bundle API: wrap workflows into deployable bundles
- Ingress service: SSE/REST/Socket.IO + gateway checks
- Processor service: queue‑driven execution of bundles
- Metrics service: aggregated stats for autoscaling & ops
- Storage: Postgres + Redis + object store integration
KDCube’s ReAct v2 agent is timeline‑first: every turn event is captured as structured blocks onto turn timeline
that become the source of truth for memory, artifacts, and future reasoning.
Turn timeline evolves into a running conversation timeline.
This is not a thin wrapper around tool calls — it’s a full stateful operating layer.
Highlights
- Timeline as ground truth: user prompts, tool calls/results, artifacts, and decisions are stored as blocks.
- Rendering: timeline can be rendered differently to an agent based on filters / cache TTL / compaction state.
- Compaction + cache checkpoints: stable prefixes + safe tail edits at scale.
- Source pools + citations: Perplexity‑style traceability with stable source IDs.
- Artifact paths & rehydration:
fi:/ar:/so:/tc:logical paths + rehosting on demand. - Tool‑aware UX: widgets stream into timeline banners (web search, fetch, exec, panels).
- ReAct tools:
react.read,react.write,react.hide,react.memsearch,react.patchto recover or reshape context. - Turn snapshots & versioning: each turn persists a timeline snapshot + data snapshot; edits produce a new version in the current turn namespace, making state recoverable and replayable.
The timeline is temporal and single‑source‑of‑truth:
- It powers UI reconstruction (user messages, attachments, artifacts created, sources used, canvas streams, thinking blocks).
- It powers agent context rendering (the model sees the same ordered timeline, filtered by policy).
- The agent can reshape the tail (e.g., hide large blocks) to keep context clean and efficient.
- Cache‑aware visibility: cache TTL is tracked so older context can be briefly surfaced (not compacted away) with re‑read paths for recovery.
- Announce system: ephemeral signals to the agent (recent pruning, active plans, important memories, current sources pool).
Timeline sketch (schematic):
graph LR
H[History blocks] --> CP1[CP prev]
CP1 --> C[Current turn blocks]
C --> CP2["CP pre‑tail"]
CP2 --> CP3[CP tail]
CP3 --> SP[Sources pool]
SP --> A[Announce]
R["react.read(fi:/ar:/so:/tc:)"] -.-> H
HIDE["react.hide(path)"] -.-> CP3
At a glance (timeline‑first loop):
graph LR
U[User + Attachments] --> TL[Timeline Blocks]
T[Tools / Exec] --> TL
TL --> SP[Sources Pool]
TL --> R[Render + Cache/Compaction]
R --> D[ReAct Decision]
D --> T
TL --> UI[Timeline UI / Widgets]
graph TD
UI[Web UI / Client] -->|HTTPS| NGINX[Web Proxy]
AUTH[Delegated Auth / SSO] -->|token exchange| NGINX
NGINX -->|SSE/REST| INGRESS[Chat Ingress]
INGRESS -->|enqueue| Q[Redis Queues]
Q --> PROC[Chat Processor]
PROC --> BUNDLES[Bundles / Workflows]
BUNDLES -->|events| RELAY[ChatRelay + Redis PubSub]
RELAY -->|fan-out| INGRESS
BUNDLES --> CTX[Context & Memory]
CTX --> PG[(Postgres)]
CTX --> S3[(Object Storage)]
PROC --> EXEC[Isolated Exec]
PROC --> TOOLS[External Tools/APIs]
INGRESS --> METRICS[Metrics Service]
PROC --> METRICS
classDef infra fill:#f7f2ff,stroke:#b69ad6,color:#2b1b4f;
classDef aws fill:#e8f4ff,stroke:#7aa7d6,color:#0b2b4f;
class INGRESS,PROC,RELAY,METRICS infra;
class PG,S3 aws;
- ECS deployment with proper autoscaling (in progress)
- Copilot‑style workspace UX (new timeline/announce events/workspace organization)
- Modeling external processes and ontologies with tools/skills/runtime and ReAct
Planned deployment options (next steps):
- AWS ECS/Fargate (first‑class)
- Kubernetes (EKS / GKE / AKS) Docker Compose is already supported for local and small‑scale (with EC2) setups.
- ReAct v2 + Timeline UX:
app/ai-app/docs/sdk/agents/react/turn-data-README.md - Streaming protocol & SSE events:
app/ai-app/docs/clients/sse-events-README.md - Bundles (multi‑workflow hosting on shared capacity):
app/ai-app/docs/sdk/bundle/bundle-README.md - Economics & accounting:
app/ai-app/docs/sdk/infra/economics - Monitoring & autoscaling metrics:
app/ai-app/docs/service/README-monitoring-observability.md
- Architecture (short):
app/ai-app/docs/arch/architecture-short.md - Gateway config & ops:
app/ai-app/docs/service/gateway-README.md - SDK index:
app/ai-app/docs/sdk
- Smart timeline + compaction:
app/ai-app/docs/sdk/agents/react/turn-data-README.md - Tooling + isolated runtime (tools inside code):
app/ai-app/docs/sdk/runtime - Attachments, artifacts, and traceability:
app/ai-app/docs/sdk
We’re actively looking for collaborators and early adopters. If you’re building AI assistants or copilots and want to ship fast with control over runtime, tooling, and costs—KDCube is for you.
Project site: https://kdcube.tech/