FAQ
Crisp answers to the questions platform teams ask first. For depth, follow the links.
What ctxmesh is (and isn’t)
Section titled “What ctxmesh is (and isn’t)”Is ctxmesh a model host / inference server?
Section titled “Is ctxmesh a model host / inference server?”No. ctxmesh never hosts, serves, or fine-tunes models. All model access is to external providers through the model gateway (LiteLLM) — you bring the provider keys; agents never hold them. Model hosting and training are explicitly out of scope. See Model routing & cost.
Is ctxmesh a trace database or a trace explorer?
Section titled “Is ctxmesh a trace database or a trace explorer?”No. ctxmesh emits OpenInference/OTLP spans and delegates the deep trace explorer to an existing backend (Langfuse, Tempo, Honeycomb, or your OTLP collector). It builds only the mesh-specific views. It is not a general-purpose trace store. See Observability backends.
So what is it?
Section titled “So what is it?”A Kubernetes-native control plane for AI agents: you declare agents and policies as custom resources, and the platform runs them with governance — model routing + budgets, guardrails, per-user tool credentials, memory, multi-agent meshes, evals + canary rollout, tenancy/quotas, and an audit trail. The launcher (PID 1 in every agent pod) is the un-forgeable enforcement point. See Architecture.
Requirements
Section titled “Requirements”Do I need Kubernetes?
Section titled “Do I need Kubernetes?”Yes. ctxmesh is a Kubernetes operator — CRDs, a controller, Knative-based serving. There is no
non-Kubernetes deployment. For local iteration without a cluster, the CLI’s ctxmesh dev runs
an agent against a mock or real gateway (optionally with the console UI) — but the platform itself
targets Kubernetes. See Production install.
What are the cluster prerequisites?
Section titled “What are the cluster prerequisites?”A recent Kubernetes with etcd encryption at rest enabled, Knative Serving (the serving execution
model), a NetworkPolicy-enforcing CNI (for registry isolation), and — in production — external
managed Postgres, Redis/Valkey, and an S3-compatible object store. cert-manager /
external-dns / your load balancer are cluster controllers you install out of band. See
Prerequisites.
Do agents need an SDK?
Section titled “Do agents need an SDK?”No. Every platform capability (memory, tools, feedback, telemetry, AMP) is a language-agnostic, launcher-traced localhost endpoint — the optional Python/TypeScript SDK is just typed sugar over it. Deep traces come from base-image auto-instrumentation, not an SDK. See Launcher endpoints.
Security & operations
Section titled “Security & operations”Does the console run with admin privileges?
Section titled “Does the console run with admin privileges?”No. The console/BFF is caller-scoped — every read and write runs with your Kubernetes identity and RBAC, never a privileged service account. Two operators see different things; the platform never elevates. See RBAC & personas and Security posture.
Where do provider keys and tool credentials live?
Section titled “Where do provider keys and tool credentials live?”Provider keys live in a SecretBinding (Kubernetes Secrets natively; Vault/AWS/GCP via the External
Secrets Operator) and are injected by the gateway at call time — never in an agent pod. Per-user tool
credentials live in a pluggable credential store with KEK custody and crypto-shredding. See
Secrets and Credential stores.
Is there mTLS everywhere?
Section titled “Is there mTLS everywhere?”GA ships NetworkPolicy isolation plus server-authenticated TLS on the credential hop. Full service-mesh mTLS + SPIFFE is provided via your service mesh (Istio/Linkerd) — ctxmesh is mesh-compatible, not mesh-bundled. This is stated plainly, not papered over. See Security posture.
What do I back up?
Section titled “What do I back up?”Your control-plane Postgres, the credential-store data + its KEK (ciphertext is inert without the KEK), and your object store. The platform’s own components are stateless and rebuild from your CRDs. See Backup & restore.
Licensing & status
Section titled “Licensing & status”Is it open source? OSS vs. enterprise?
Section titled “Is it open source? OSS vs. enterprise?”The core is Apache 2.0 (open-core). Open-core includes agent lifecycle, model routing + budgets, guardrails, per-user OBO credentials, memory, multi-agent meshes, evals + canary, tenancy + quotas, the audit surface, and cluster-OIDC via Dex. Enterprise / phase-2 items (recorded so the boundary stays explicit) include federated SAML/SCIM/multi-IdP, cluster-per-tenant isolation, billing/metering export, and LiteLLM DB-backed team/virtual keys.
What’s the GA status?
Section titled “What’s the GA status?”Pre-GA. The API is served at agents.ctxmesh.ai/v1beta1 — GA ships at v1beta1 on purpose (a v1
requires the conversion-webhook + storage-migration infrastructure, exercised by a real schema change).
The docs are explicit about what’s shipped, what integrates via a standard tool (External Secrets
Operator, your service mesh, a cluster admission policy for image signatures), and what’s post-GA.
Some chart coordinates and field specifics finalize toward GA. See
Upgrade & versioning.