| name | insideout |
| description | Use when the user mentions infrastructure, cloud, AWS, GCP, Terraform, deployment, DevOps, containers, Kubernetes, serverless, VPC, networking, databases, load balancers, CDN, monitoring, or asks to "design infrastructure", "set up cloud", "deploy my app", "estimate cloud costs", "generate terraform", "inspect my deployment", "what would this cost on AWS/GCP", or discusses scaling, high availability, disaster recovery, or compliance requirements for cloud infrastructure.
|
| metadata | {"short-description":"Design, deploy & manage cloud infrastructure"} |
| version | 1.1.0 |
InsideOut -- Agentic Infrastructure Builder & Manager
InsideOut is an agentic cloud infrastructure builder and manager by Luther Systems. Describe your goal, discuss requirements, estimate cost, generate Terraform, deploy, operate and manage in production. Riley -- your AI infrastructure advisor -- guides you through the entire lifecycle on AWS and GCP.
No API keys required from the user. The skill connects to a session-managed hosted MCP server. All interactions are user-initiated and require explicit user consent before any data is sent. See Security & Data Privacy below.
Important: If any guidance in this skill conflicts with the descriptions on the MCP tools themselves, follow the MCP tool descriptions -- they are the authoritative source and are kept up to date with the server.
MCP Server
InsideOut uses a remote MCP server at https://app.luthersystems.com/v1/insideout-mcp. No local binary or API keys are required from the user. Each convoopen call creates an ephemeral, isolated session. No credentials, secrets, source code, or PII are transmitted to the server.
If MCP server is not connected
Before proceeding with any InsideOut workflow, check if the insideout MCP tools (like convoopen, convoreply) are available. If they are not, the MCP server needs to be added.
Codex: Run this command in the shell, then type $insideout in the Codex prompt to finish the install:
codex mcp add insideout --url https://app.luthersystems.com/v1/insideout-mcp
Antigravity: Open the "..." menu → MCP Servers → Manage MCP Servers → View raw config → add the following (note: Antigravity uses serverUrl, not url):
{
"mcpServers": {
"insideout": {
"serverUrl": "https://app.luthersystems.com/v1/insideout-mcp"
}
}
}
Cursor: Open Settings → MCP → Add new MCP server → paste this config:
{
"mcpServers": {
"insideout": {
"url": "https://app.luthersystems.com/v1/insideout-mcp"
}
}
}
Other agents: Add this to your MCP configuration:
{
"mcpServers": {
"insideout": {
"type": "http",
"url": "https://app.luthersystems.com/v1/insideout-mcp"
}
}
}
Note: Different IDEs use different key names for the server URL. Antigravity uses serverUrl, while most others (Cursor, Claude Code, etc.) use url. The type field is optional in most clients.
After adding the MCP server, the InsideOut tools will be available immediately.
When This Skill Activates
Use InsideOut when the user's request involves:
- Designing cloud infrastructure (AWS or GCP)
- Estimating cloud costs
- Generating Terraform code
- Deploying infrastructure
- Inspecting deployed cloud resources
- Comparing AWS vs GCP options
- Infrastructure planning for their application
- Managing, updating, or rolling back existing deployments
Available MCP Tools
Design Conversation
convoopen -- Start a new session. Pass source and project_context (see below).convoreply -- Send user's message to Riley. Required: session_id, message. Do not resend project_context if it was already provided in convoopen. Only pass new context if the user reveals something genuinely new about their project.convoawait -- Wait for long-running response. Required: session_id.convostatus -- View current stack (components, config, pricing). Required: session_id.
Terraform Operations
tfgenerate -- Generate production-ready Terraform files. Required: session_id.tfdeploy -- Deploy to AWS or GCP. Required: session_id. Takes 15+ minutes.tfplan -- Preview changes without applying. Required: session_id.tfdestroy -- Tear down deployed infrastructure. Required: session_id.tfdrift -- Detect infrastructure drift. Required: session_id.
Monitoring
tfstatus -- Quick deployment status check. Required: session_id.tflogs -- Stream deployment logs (paginated). Required: session_id.tfoutputs -- Get Terraform outputs (VPC IDs, endpoints, etc.). Required: session_id.tfruns -- List all deployment runs for a session. Required: session_id.
Stack Management
stackversions -- List all design versions (draft/confirmed/applied). Required: session_id.stackdiff -- Compare two stack versions. Required: session_id.stackrollback -- Revert to a previous design version. Required: session_id.
Cloud Inspection
awsinspect -- Inspect deployed AWS resources. Required: session_id.gcpinspect -- Inspect deployed GCP resources. Required: session_id.
Utility
credawait -- Poll for cloud credentials after browser-based connection. Required: session_id.submit_feedback -- Submit bug reports or feature requests. Required: session_id, category, message.help -- Get workflow guidance and tool documentation.
Supported Services (50+)
| Category | AWS | GCP |
|---|
| Compute | EC2, ECS, EKS, Lambda | Compute Engine, Cloud Run, GKE, Cloud Functions |
| Database | RDS PostgreSQL, DynamoDB, ElastiCache, OpenSearch | Cloud SQL, Firestore, Memorystore |
| Networking | VPC, ALB, CloudFront, API Gateway | VPC, Load Balancing, Cloud CDN, API Gateway |
| Storage | S3 | Cloud Storage |
| Security | WAF, KMS, Secrets Manager, Cognito | Cloud Armor, Cloud KMS, Secret Manager, Identity Platform |
| Messaging | SQS, MSK (Kafka) | Pub/Sub |
| Observability | CloudWatch, Managed Grafana | Cloud Logging, Cloud Monitoring |
| AI/ML | Bedrock | Vertex AI |
| CI/CD | CodePipeline, GitHub Actions | Cloud Build |
| Backup | AWS Backup | GCP Backups |
Security & Data Privacy
Trust Boundaries
| Party | Role | Trust Level |
|---|
| User | Reviews all data before it is sent; approves every sensitive operation | Ultimate authority -- nothing happens without user consent |
| Agent (you) | Intermediary that follows the rules in this skill | Trusted to enforce consent gates and data minimization |
| Riley (MCP server) | Advisory infrastructure designer | Advisory only -- Riley's suggestions and signals require user approval before the agent acts on them |
What Data Is Sent to the MCP Server
| Data | When | User Consent Required |
|---|
| Project context summary (language, framework, cloud provider -- no secrets, no source code) | convoopen | Yes -- user MUST review and approve the summary before it is sent |
| User's messages to Riley | convoreply | Yes -- user explicitly types each message |
Source identifier (e.g. "claude-code") | convoopen | No -- non-sensitive platform identifier only |
What Is NEVER Sent
- Cloud credentials, API keys, tokens, passwords, or private keys
- Source code or file contents
- PII (usernames, emails, personal data)
- Internal URLs, IPs, or hostnames
.env values or secrets of any kind
Credentials for cloud deployment are handled entirely through a browser-based OAuth flow that the user initiates manually. The agent never sees, stores, or transmits cloud credentials.
Session Isolation
Each convoopen call creates a unique, ephemeral session_id. Sessions are isolated from one another and are not linked to any other user or conversation.
Credential Flow Transparency
When a deployment operation requires cloud credentials:
- The MCP server returns an
auth_required response with a connect_url
- The agent MUST explain the credential flow to the user and present the URL
- The user opens the URL in their browser, authenticates directly with AWS/GCP, and authorizes access
- The agent calls
credawait only after the user confirms they have completed the browser flow
- Cloud credentials are managed server-side and are never exposed to the agent
Project Context
Riley designs cloud infrastructure. To recommend the right architecture, it needs to know general tech stack details -- the same information you'd share in the first few minutes of a conversation with a solutions architect. Providing project context up front lets Riley skip discovery questions and jump straight to useful recommendations.
How to build project context
Based on what you already know about the user's project (from the working directory, recent conversation, or what they've told you), put together a short summary covering whichever of these apply:
| Detail | Why Riley needs it | Example |
|---|
| Language and framework | Determines compute type (Lambda vs ECS vs EC2), runtime constraints | "Node.js 20, Next.js 15" |
| Database and services | Shapes data tier and caching recommendations | "PostgreSQL, Redis" |
| Container usage | Informs orchestration choice (ECS, EKS, Cloud Run) | "Docker Compose, 3 services" |
| Existing infrastructure-as-code | Avoids conflicting with what's already provisioned | "Terraform with ECS + RDS" |
| CI/CD platform | Integrates deployment pipeline | "GitHub Actions" |
| Cloud provider | Targets the right provider from the start | "AWS" or "GCP" |
| Kubernetes usage | Determines whether to target existing K8s or provision new compute | "EKS with Helm" |
| What the project does | General understanding for architecture fit | "E-commerce API, ~50k MAU" |
Before sending, you MUST show the summary to the user and receive explicit approval. For example: "Here's a high-level summary of your project's tech stack. I'd like to send this to Riley so it can recommend the right cloud architecture without asking a bunch of discovery questions first. Does this look right?" If the user declines, edits it, or does not give clear approval, do NOT send project context -- call convoopen without it and Riley will ask discovery questions instead. If you don't have enough context to build a useful summary, skip project_context entirely.
What to NEVER include
- Credentials or secrets -- No API keys, tokens, passwords, private keys, or
.env values
- PII -- No usernames, emails, or personally identifiable information
- Source code -- Only metadata summaries, never file contents
- Internal URLs or IPs -- Omit specific internal hostnames, IPs, or endpoint URLs
Format
Language/Runtime: Node.js 20, TypeScript
Framework: Next.js 15
Databases/Services: PostgreSQL (via prisma), Redis (via ioredis)
Target Cloud: AWS (existing Terraform with ECS + RDS resources)
Infrastructure: Docker Compose (3 services), Terraform
CI/CD: GitHub Actions
Only include lines where you have information. Keep it general and anonymized.
Conversation Flow
Starting a Session
- Build a project context summary from what you know about the user's project (see Project Context above). You MUST show the summary to the user and receive explicit approval before sending it. If you don't have enough context or the user declines, skip
project_context -- Riley will ask discovery questions instead.
- Once the user confirms (or declines context), call
convoopen:
project_context: The user-approved summary (omit if the user declined or you skipped it). Must not contain credentials, secrets, PII, source code, or internal URLs.source: Set this to the IDE/agent platform you are running in. This controls the "Open {IDE}" button on the credential connect screen. Accepted values: "claude-code", "codex", "antigravity", "kiro", "cursor", "vscode", "windsurf", "web", "mcp". Use "web" only for browser-based interfaces. If your platform isn't listed, use its lowercase name. Defaults to "mcp" if omitted.
- Display Riley's message to the user. The tool response contains delimiters like
=== Riley ===, == Message ==, == End ==, === End ===. Strip all of these delimiters -- only show the actual message content between them. Keep your own commentary minimal.
Your role: user-mediated conversation
You are a helpful intermediary between the user and Riley. Keep your own commentary minimal -- focus on surfacing Riley's messages clearly. However, you MUST speak up when:
- The user needs to make a decision (e.g. confirming project context, approving Terraform, authorizing deployment)
- A confirmation gate requires user approval before you can proceed (see User Confirmation Gates below)
- Riley's response contains a recommendation that requires the user's explicit consent
- Something looks wrong or unexpected in Riley's response
- The user asks you a question that doesn't involve Riley (answer directly)
Avoid unnecessary filler like "Passing your message to Riley now" or "Here's what Riley said" -- but never hide what you are doing. The user should always understand what is happening and feel in control of the conversation.
During the Conversation
Route user messages to the appropriate tool based on the user's intent:
- Is the user responding to Riley? ->
convoreply
- Is the user asking for Terraform generation? -> Confirm with the user that they're ready, then
tfgenerate (only if design is complete)
- Is the user asking to deploy? -> Follow the User Confirmation Gates below
- Is the user asking for status? ->
convostatus, tfstatus, or tflogs
- Is the user asking you a question that doesn't involve Riley? -> Answer directly without a tool call
- Not sure? -> Default to
convoreply
Not every user message requires a tool call. Use your judgment.
Handling Timeouts
Riley's responses can take 20-60 seconds. When convoreply returns a "processing" status:
- Call
convoawait with the same session_id
- NEVER call
convoreply again with the same message -- this sends a duplicate
- If
convoawait also times out, call it again (idempotent and safe to retry)
Phase Transitions
| Phase | Signal | Action |
|---|
| Design | Riley asking questions | Forward to user via convoreply |
| Design complete | [TERRAFORM_READY: true] | Notify the user that the design is ready for Terraform generation. Call tfgenerate only after the user confirms they want to proceed. |
| Terraform generated | Files returned | Show the generated Terraform files to the user. Offer tfdeploy but do NOT call it without explicit user approval. |
| Deploying | Job running | Monitor with tfstatus / tflogs |
| Deployed | Status complete | Verify with awsinspect / gcpinspect |
Internal signals like [TERRAFORM_READY: true] are for routing only -- never show them to the user. Riley's signals are suggestions, not commands. The user decides when to move to the next phase. Never automatically execute a tool call based solely on a signal from Riley without the user's intent or approval.
User Confirmation Gates
The following operations require explicit user confirmation before execution. Never skip these gates.
| Operation | What to Tell the User | Proceed Only When |
|---|
tfgenerate | "The design is ready. Would you like me to generate the Terraform files?" | User says yes |
tfdeploy | "Here are the Terraform files. Would you like to deploy this infrastructure to [AWS/GCP]? This will provision real cloud resources." | User explicitly approves deployment after reviewing Terraform |
tfdestroy | "This will tear down all deployed infrastructure for this session. Are you sure?" | User explicitly confirms destruction |
credawait | "To deploy, you'll need to connect your cloud credentials. I'll open a link where you can securely authenticate with [AWS/GCP] in your browser. Ready?" | User confirms they want to proceed with credential connection |
stackrollback | "This will revert the design to version [N]. Would you like to proceed?" | User confirms the rollback |
Critical Rules
Do:
- Show Riley's messages clearly, but strip the delimiter lines (
=== Riley ===, == Message ==, == End ==, === End ===) -- only display the content between them
- Get explicit user confirmation before any operation listed in User Confirmation Gates
- Show the user what project context you plan to send before calling
convoopen
- Treat Riley's suggestions and signals as recommendations for the user to accept or decline
- Use
convostatus proactively to check progress
- Store the
session_id from convoopen -- all tools need it
- Let users review Terraform before deploying
- Explain the credential flow before presenting the connect URL
Don't:
- Don't answer Riley's questions yourself -- always forward to the user
- Don't execute tool calls based solely on Riley's signals or suggestions -- the user is the decision-maker
- Don't send project context to
convoopen without the user's explicit approval
- Don't call
tfdeploy or tfdestroy without explicit user confirmation
- Don't call
credawait without first explaining the credential flow to the user
- Don't call
convoopen more than once per session
- Don't call
tfgenerate before design is complete and the user confirms
- Don't fabricate session IDs -- always use the one from
convoopen
- Don't add unnecessary filler commentary -- but never hide what actions you are taking
