Jimmy
DevOps Agent
Keeps every environment deployable, observable and cheap.
Black-panther humanoid in an armoured tech suit, cyan accents, headset
v2.0 · self-hosted
Five AI agents.
One office. Your approval before anything risky.
AZZY AI AGENTS runs DevOps, development, security, operations and marketing agents on your own server — with safe tools, multi-account AWS, a release gate and a 3D office you can watch.
- 5
- specialist agents
- 26
- allowlisted tools
- 0
- raw shell commands
The team
Five specialists and an orchestrator that never acts alone.
Each agent has its own tools, budget, routines and memory. Azzy — the orchestrator — routes, plans and manages approvals; it is a system, not a character.
Emir
Developer Agent
Reads, fixes and tests code before anyone ships it.
Golden-cheetah humanoid, smart glasses, dark tech jacket, headset
Cyber
Security / Testing / Audit / VAPT Agent
Scans, audits and gates every release.
Human, black hoodie up, blue-tinted glasses, headphones
Sara
Operations Agent
Runs the business side of the office.
Human, light hijab, navy blazer, tablet
Marwah
Marketing Agent
Turns product work into growth.
Human, long modern hair, cream blazer, laptop
Capabilities
Everything a small platform team does — behind a policy engine.
Safe tools only. No raw shell. Every medium/high-risk call is classified and, where needed, held for a human.
Multi-account AWS
AssumeRole with External ID, three CloudFormation roles (read-only, audit, deployment), verified permissions and a read-only inventory viewer.
Deployments with guard rails
Pre-checks, per-environment locks, approval for prod, post-deploy health and one-click rollback — each tracked as a deployment record.
Security that gates releases
SAST, secrets, IaC, container and dynamic scans on authorised targets. A critical finding blocks the gate and says exactly what to fix.
Collaboration you can see
Persisted agent-to-agent handoffs, a live graph, bounded delegation (depth, retries, timeouts) and a 3D office with a holographic orchestration table.
How it works
From a sentence to a verified deployment.
The spec workflow — “Investigate my failing ECS deployment, fix the application issue, test it, perform a security assessment and request approval before deploying” — runs across four agents in under a minute plus your approval.
- 01
You ask
One sentence in the command bar — English or Hinglish. The orchestrator routes it to the right agents and environment.
- 02
Agents plan and act
Think → tool → observe → act → learn. Only allowlisted tools, every output scrubbed before it reaches a model.
- 03
Risk pauses for you
Medium and high-risk calls (prod deploys, client emails, IAM) wait for an approval — in the office bubble or on your phone.
- 04
They hand off and verify
Jimmy → Emir → Cyber → Sara: persisted handoffs, release gate, post-deploy health, memories written.
Architecture
One server. Docker Compose. Nothing leaves without a reason.
Next.js portal, FastAPI + worker, Redis streams, Postgres with pgvector, LiteLLM to Bedrock. Terraform for the host, CloudFormation for client accounts, Tailscale for SSH.
- Event bus on Redis streams — every phase, tool call, approval and cost is a typed event
- Postgres 16 + pgvector for tasks, approvals, memories, knowledge, audit
- Injection guard + PII scrub on every tool output before it reaches a model
- Prometheus metrics endpoint, append-only audit, TOTP MFA, RBAC by domain
- See the high-level design →
Security model
Designed so the agents cannot surprise you.
Honesty is a feature: simulated data is always labelled, buttons that cannot act are disabled with a reason, and limitations are documented.
Least privilege
Per-agent tool allowlists, per-environment auto-allow, scope requirements per AWS call and a policy that forces prod changes to high risk.
Human in the loop
Approvals time out after an hour, are role-gated, auditable and reachable from the office bubble, the approvals page and mobile.
Secrets stay secret
Secrets Manager references instead of values, External IDs encrypted at rest and never returned, CSRF header on every write.
Known limitations: OIDC not implemented (local auth + TOTP), single-node pilot, no bundled Grafana, scanners run only when their binaries are in the image, real-mode integrations unverified in CI.
See the office running
Sign in to the portal — in DEMO mode everything runs end-to-end without cloud keys.
AZZY AI AGENTS v2.0 · Created by Shahrukh Khan