Restack
Type: full-code · Vendor: Restack · Language: Python, TypeScript · Status: active · Status in practice: emerging · First released: 2024-07-27
Backend platform for building long-running AI agents on top of Temporal and Kubernetes, with workflows-as-code, MCP-exposed tools, ClickHouse-backed context, and a product-team-facing visual interface.
Description. Restack is a vendor-managed agent platform for enterprise teams. Engineers write workflows in Python or TypeScript (`@restackio/ai`) by composing functions called via a `step` primitive; the platform runs these on a Temporal-backed durable engine deployed on Kubernetes so agents can run for weeks or months. Python functions are exposed to the agent as MCP tools, long-term context lives in a ClickHouse store, and product teams manage prompts, versions, and tests through a UI.
Agent loop shape. Restack agents and workflows are durable Temporal-backed processes. A workflow is a multi-step function that uses the `step` primitive to call user-defined functions (the activities); the framework records each step so that on crash or restart the run resumes from the last committed step. Agents are long-running variants that can wait on events, call MCP tools, pull context from the ClickHouse store, and be scheduled via the platform's Schedules feature. Restack manages the underlying Kubernetes + Temporal stack so users do not run Temporal directly.
Primary use cases
- long-running enterprise AI agents (weeks-to-months horizons)
- durable workflows composed from Python or TypeScript functions
- MCP-tool integrations exposing in-house Python code
- self-hosted agent deployments on AWS / GCP / Azure / on-prem
Key concepts
- Workflow → durable-workflow-snapshot (docs) — Multi-step orchestration coordinating model calls, API requests, and logic with durable state.
- Step / Function → tool-use (docs) — `step()` primitive inside a workflow calls user-defined functions (the activities of the run).
- Agent → agent-resumption (docs) — Long-running variant of a workflow that handles errors, retries, and interruptions automatically.
- Context store (docs) — Built-in long-term memory powered by ClickHouse; stores past interactions and retrievable context.
- MCP-out-of-the-box → mcp (docs) — Python functions are exposed to the agent as MCP tools by default.
- Schedules → scheduled-agent (docs) — Cron / calendar schedules on workflows, agents, and functions.
Patterns this full-code implements —
- ★★Agent Resumption
Durable Execution lets Restack agents run for months, handling errors, retries, and interruptions automatically.
- ★★Scheduled Agent
Restack has a dedicated Schedules feature for workflows, agents, and functions; workflows can also be scheduled programmatically via `scheduleWorkflow`.
- ★★Tool Use
MCP is the default tool surface; Python functions are exposed as tools the agent can call.
- ★★Model Context Protocol
MCP is the integration mechanism: Restack uses MCP to expose Python functions as agent tools.
- ★Durable Workflow Snapshot
Built on Temporal and Kubernetes; Durable Execution is the marquee infrastructure property.
- ★★Event-Driven Agent
Workflows can be scheduled from agents or other workflows; Restack exposes webhooks. No verbatim quote on event-driven agent triggers in the pages successfully fetched; honest downgrade from first-cl…
- ★★Step Budget
Workflows compose via the `step` primitive but no explicit step-budget cap is documented in the pages fetched; honest downgrade from first-class.
- ★★Supervisor
Workflows orchestrate functions and can call sub-workflows from an agent. No explicit router/supervisor primitive equivalent to AgentKit's was found on fetched pages; honest downgrade from first-clas…
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Alternatives & relatives
Listed as alternative by (3)
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: partial