The Cloud Clockwork is a cloud and AI platform consultancy. We engineer cutting-edge, production-grade agent platforms on AWS and Kubernetes — model gateways, policy-as-code governance, and full-stack observability — so your teams ship AI products, not plumbing.
CAPABILITIES
Scalable, resilient, cost-optimized infrastructure on AWS, Azure, and GCP. Multi-account, least-privilege, Terraform end to end.
Enterprise agent platforms: model routing, MCP tool gateways, policy-as-code governance, and full-stack observability.
GitOps with Helm and ArgoCD, CI/CD on every push, Kubernetes-native autoscaling, and reproducible infrastructure.
OPEN-SOURCE · AWS-NATIVE AGENT PLATFORM
A domain-agnostic, cloud-native agent platform on AWS Bedrock AgentCore and the Strands SDK. A team writes declarative blueprints — one YAML per agent, a prompt, a five-line handler — and gets back a production agent fleet: microVM-isolated runtimes, a federated tool gateway, policy-as-code authorization, pluggable observability, and least-privilege IAM, all provisioned by Terraform that reads the same YAML the runtime reads. Decoupled at every layer that usually locks you in.
Validated in production, not a demo. The platform runs QITP, a regulated, autonomous long-only investing system: it sizes risk-bounded bracket orders behind a deterministic risk engine and a 2FA gate and places them on Interactive Brokers (paper trading, rolling to live), with nine agents, nine MCP servers, and a Step Functions pipeline, all as declarative layers on the same engine.
Open source, MIT-licensed. Clone it, drop in your own blueprints, and stand up your own production agent fleet on AWS. Inspect the Terraform, the Strands runtime adapters, and the AgentCore and gateway wiring, or read the full architecture docs.
The reusable engine. Declarative blueprints in, a secure microVM agent fleet out, on any domain and any model.
QITP: a regulated, autonomous AWS-native trading system running on the exact same platform. Proof, not a separate codebase.
A single YAML provisions both the live runtime and its Terraform infrastructure. The pod and the infra cannot drift — both are projections of one source of truth.
Agents and MCP tool servers are the same deployable: ARM64 / Graviton microVMs on identical CI, IAM, and observability rails, under one audit trail.
One field swaps the model: Bedrock, 100+ models behind a LiteLLM gateway, Anthropic, or Vertex. Zero code, zero Terraform change.
Agents know one endpoint. The AgentCore Gateway translates MCP to Lambda, OAuth2 MCP runtimes, and OpenAPI. New tools reach the whole fleet with no rebuild.
Cedar authorization at the gateway, default-deny, authored in plain English and compiled to policy. Every version hashed and auditable.
AWS X-Ray and Langfuse at once, plus an LLM-as-judge evaluation harness with online sampling and per-trace token and cost tracking.
Drop a blueprint, get a fleet. A for_each over the YAML provisions ECR, CodeBuild, least-privilege IAM, the runtime, and gateway targets — fully diffable in a PR.
Runs a regulated, money-touching autonomous system that places autonomous bracket orders on a connected Interactive Brokers account (paper trading, rolling to live) — nine agents, nine MCP servers, and a deterministic risk engine, all as declarative layers.
AWS ARCHITECTURE
SELF-HOSTED CONTROL PLANE
An agentic platform built for high-performance solopreneurs and small teams. Production-grade ecosystem running 24/7 on owned hardware, behind infrastructure and security under full control. AgentiCore agents coordinate autonomously to manage infrastructure, create content, and execute workflows.
AgentiCore multi-agent runtime with 7 specialists — diagram, video editor, NotebookLM, FinOps, publisher, Anton infra agent, and brain keeper. Agent-to-agent dispatch via AgentiBridge, fleet-wide memory, and an ops oracle for self-healing.
LiteLLM proxy with multi-provider routing, 18 MCP servers across 8 tool groups (500+ tools), private AI chat UI, and model-as-agent architecture.
NotebookLM integration, AI video editor pipeline, Gemini + DALL·E media generation, automated publishing, and Paper2Slides conversion.
Grafana + OpenTelemetry + ClickHouse stack, Langfuse tracing and cost tracking, SonarQube quality gates — full visibility across the agent fleet.
Cloudflare Zero Trust perimeter, OpenBao secrets management, External Secrets Operator for K8s, and OAuth shim for secure MCP access.
Self-hosted on Unraid with a 4-node K3s cluster, ArgoCD GitOps, self-hosted GitHub Actions runners, Traefik ingress, and Cloudflare Tunnel exposure.
PostgreSQL, Redis Streams, CouchDB, MongoDB, ClickHouse, S3 artifact storage with Drive sync, Obsidian semantic vault, and a real-time event bus.
Jira + Confluence, Plane, Telegram bot, Google Drive, GitHub MCP, and router management driven from the AI layer.
CI/CD on every push, Helm-based deployments, dev/prod parity, and automated ArgoCD image updates — sub-2-minute deploy pipeline.
OPEN SOURCE
An open-source control plane that transforms solitary AI coding sessions into coordinated agent workforces. Self-hosted, MIT-licensed, zero vendor lock-in.
Workforce network — semantic search across sessions, agent-to-agent dispatch, cross-project handoff, and MCP protocol front door.
Execution engine — containerized task dispatch via KEDA, OpenAI-compatible endpoints, Kubernetes-native autoscaling from 1 to 1000 containers.
In-agent harness — context compression, retry circuit breakers, broadcast synchronization, attention refresh, and cost-saving guardrails.
Memory vault — Obsidian-compatible markdown storage, arc-based focus blocks with heat scoring, synthesis engine, and fleet-wide signal broadcasting.
MEMORY ARCHITECTURE
A persistent memory system modeled after the human brain — not as a metaphor, as an architecture. Knowledge enters through two write paths, gets clustered into arcs, scored by a heat engine, and self-organizes across hemispheres. Every two hours the tick reconciles the vault and broadcasts fleet-wide awareness. No manual curation. It never forgets.
