Skip to content

Wiki

One page per subject. Try the demo (a beginner's walkthrough on a laptop) and Deploy your first group are the short paths; these are the full ones. Roles and the threat model sit beside them.

Page What it covers
Deploy on GCP Terraform, images, Helm, the first zone and group, what the console's service account may do, day 2, teardown
Deploy on AWS The same on EKS with DynamoDB, S3, Secrets Manager and a load balancer, and what its IAM role may do
Deploy on Azure AKS with Table Storage, Blob, Key Vault, Envoy Gateway and Let's Encrypt; subscriptions without availability zones
IAM and network on GCP, AWS, Azure Every identity Ramen uses on each cloud and what it may do — console, nodes, workers, per-group build identities, the evals Jobs, Workload Identity / IRSA / federated credentials — and every network path: the edge, NetworkPolicies, firewall / security group / NSG rules, CIDRs, satellites
MCP features Every MCP capability a worker serves and how a group repo codes it, with the demo repo's examples: tools, resources, templates, prompts, tasks, MRTR, subscriptions, apps, skills, guardrails, evals, LLM configs
Groups, zones and regions The model, creating each one, deploying, scaling, rebalance, per-tool access, blocked packages, IP rules, deleting
Autoscaling Ramen scales workers between 1 and an admin's max: the decision rules, cooldowns, audit, load-test results
Sleeping workers 1.0: an idle zone scaled to 0 behind one small activator pod: what idle means, what the first caller sees, what it costs
The MCP repo The folder layout, the JSON beside each tool, resources and prompts, what a package can do since 1.0, apps and skills (1.0.0c), secrets in code, testing locally, heavy dependencies
Worker images The release image, an image built on deploy from requirements.txt (1.0), or one you pin: the builder identity per cloud, failed builds, settings
Secrets Scopes, how a value reaches your code, where values live on each cloud, rotation
Users and access Roles per group, custom roles, where a membership comes from, sign-in, service-account permissions, audit and logs
Connect a client with OAuth Registering a client; Claude Code over HTTP, the bridge for Cursor, VS Code and Claude Desktop; where the bridge keeps tokens; what goes wrong
Connect a client with a password The account side of the same flow
API keys and the API The two kinds of key, the OpenAPI page, a cheat sheet, a CI deploy
Deploy from GitHub Actions An API key for CI, the repo secret, a workflow that deploys on push and waits for the job, rotation
Throttling with Redis Per-address and per-token limits, scope and item, shared across zones
Entra ID and Google Workspace Providers on the Config page, groups read from Graph or Cloud Identity, role rules, OAuth-only roles, troubleshooting
Guardrails NeMo Guardrails rails or your own policy.py, per tool, before the call and after the result; 1.0.0c Phase 2.5: LLM-backed rails on a provider scoped to the group or the org, and the rails as a client policy the workers gate (-32029) and the bridge runs locally
Multi round-trip requests 1.0: a tool that asks the client for input and is called again: the flow, the signed state, guardrails on every round, older clients
Tasks 1.0: long-running tools that answer with a task id: who gets one, the lifecycle, the Redis a multi-worker zone needs, limits
Telemetry 1.0: OpenTelemetry export of calls and deploys, what is and is never exported, what agents are told, consent per group
Evals, LLM providers and client policy 1.0.0c: a promptfoo suite in the repo run on every deploy, report or gate per group, the LLM keys on the Config page, the evals version clients must hold, the bridge's policy bundle and editor files
SkillSpector scans 1.0.0c Phase 2: NVIDIA SkillSpector over the repo's mcp/skills/* and the served tools on every deploy, the report page, gate per group, more LLM providers, the bridge's local scan
LLM configs from the group repo 1.0.0c Phase 3: mcp/llm-config/ — client config fragments validated and versioned on deploy, accepted by every person on every sign-in path, written by the bridge, gated by the workers (-32028); the console's OTLP ingest for client telemetry
Backups What an export holds and what it does not, preview, restore, prune, reconcile
Configuration The config file, the public address (base URI) and path-prefix proxies, the state store, every variable that matters, restricting actions

Reference material sits one level deeper: transport hop by hop, the threat model, the agent skills (install them to have an agent deploy and run Ramen) and the interface contracts.