MCP server

The apps-mcp server lets coding agents (Claude Code, Codex, or any MCP client) launch and manage apps with natural language. It is a thin, well-described layer of tools, resources and prompts over the REST API — no business logic of its own, so agent launches behave exactly like UI launches.

Endpoint: https://apps.<cluster-domain>/mcp (streamable HTTP — the UI’s nginx proxies /mcp to the MCP service, so the whole platform lives on one hostname).

Connecting

The UI’s Settings page (account menu → Settings, under Connect an agent) shows your cluster’s MCP URL and the Claude Code command with copy buttons, including the sign-in client below when one is configured.

# Claude Code, on a cluster with auth on (the client id and port come from mcp.oauth)
claude mcp add --transport http --client-id nebari-apps-mcp --callback-port 33418 \
  nebari-apps https://apps.example.ai/mcp

Then run /mcp in Claude Code, select nebari-apps and choose Authenticate (or run claude mcp login nebari-apps from a shell). A browser opens on the cluster’s Keycloak; after you sign in, Claude Code keeps the token, refreshes it, and sends it with every request. On a cluster with auth off, leave out --client-id and --callback-port.

The same in a project’s .mcp.json:

{
  "mcpServers": {
    "nebari-apps": {
      "type": "http",
      "url": "https://apps.example.ai/mcp",
      "oauth": { "clientId": "nebari-apps-mcp", "callbackPort": 33418 }
    }
  }
}

Authentication

The MCP server uses standard MCP authorization: it is an OAuth 2.0 protected resource, and the cluster’s Keycloak realm is its authorization server.

1. A request to /mcp without a token gets 401 with WWW-Authenticate: Bearer resource_metadata="https://apps.example.ai/.well-known/oauth-protected-resource/mcp".
2. That document (RFC 9728) names the realm (keycloak.url + /realms/<realm>) as the authorization server.
3. The client signs the user in to Keycloak in the browser (authorization code with PKCE) and sends Authorization: Bearer <access token> on every request.

Nothing is stored per MCP session: each request carries its own token. That is what lets sessionless clients (MCP protocol 2026-07-28 and later, which Claude Code speaks) work exactly like clients that keep a session. Clients that already hold a token from this realm, such as scripts, can send the same header directly; the token must still belong to one of the pack’s clients (below).

The sign-in client

The MCP client needs a Keycloak client to sign in with. The pack cannot provision it: the nebari-operator’s NebariApp creates a confidential client, an SPA client whose redirect URIs are the UI’s own host, and a device-flow client without the authorization code flow, and none of them accepts the loopback redirect (http://localhost:<port>/callback) a CLI listens on. Keycloak’s default policy also refuses anonymous dynamic client registration, and it does not support client ID metadata documents yet, so a cluster admin creates one public client once:

# Inside the Keycloak pod (or anywhere kcadm.sh is logged in to the realm)
kcadm.sh create clients -r nebari \
  -s clientId=nebari-apps-mcp -s name="Nebari Apps MCP" \
  -s publicClient=true -s standardFlowEnabled=true \
  -s directAccessGrantsEnabled=false -s implicitFlowEnabled=false \
  -s 'redirectUris=["http://localhost:33418/callback"]' \
  -s 'attributes."pkce.code.challenge.method"=S256'

It needs the realm’s default client scopes, which include groups on a Nebari realm (group-based manager and admin checks read that claim). Then set mcp.oauth.clientId=nebari-apps-mcp (and mcp.oauth.callbackPort if you chose another port). The chart adds the client to the allowed clients of apps-api and apps-mcp, and the UI puts it in the claude mcp add command. Every user connects with the same client id and port; the port is only a local listener on the user’s machine.

Enforcement

Enforcement is layered:

• At the HTTP layer — every request to /mcp must carry a bearer token whose signature, issuer and expiry verify against the realm JWKS, and that was issued to (azp) or for (aud) one of the pack’s clients: the UI’s SPA client, the device-flow client, or mcp.oauth.clientId. A token another realm client obtained is refused with 401 (error="invalid_token"). /healthz and the metadata document need no token.
• At the MCP layer — the same check runs again before any tool call or resource read, so it also holds on transports without HTTP. Listing resources still answers, without any Apps, and prompts carry no cluster data.
• At the API layer — apps-api independently validates the JWT on every request, so the MCP can never act with more privilege than the calling user: an agent sees and manages only the apps its user owns or manages (or every app, for members of an admin group). Resources make the same API calls as the matching tools, so they are authorized the same way.

On clusters with auth disabled, /mcp needs no token, serves no metadata, and every tool works anonymously.

apps-mcp fetches the realm JWKS from keycloak.internalUrl when it is set, otherwise from keycloak.url. If that certificate comes from a private CA, every token is rejected until the CA is trusted through mcp.orgCABundle.

Tools

Tool	Purpose
describe_cluster	Source types, launchable namespaces, apps domain, the cluster’s default idle timeout and, when image is a source type, the imageRegistries images may come from — call first to pick valid options.
launch_app	Create + launch (static, Python/pixi via pixi_task, or a prebuilt container image with source_type: image, image_reference (tag or digest, from an allowed registry) and image_port, plus optional image_command, image_args and image_run_as_user; see Container images); optional resources (requests/limits, cluster defaults otherwise) and groups to restrict a private app (full Keycloak paths matched exactly, finance is /finance; a member of any listed group gets in, every other signed-in user gets 403 at the gateway); idle_timeout_minutes and keep_alive control scale to zero; manager_users/manager_groups set who else may manage it (omit to keep the current list); secrets ({NAME: value}) stores credentials in the app’s own Secret instead of env (see Secrets). Idempotent on (namespace, name) — re-launching an app you own or manage updates it instead of failing, so retries are safe; any other app is refused, never overwritten. Only the owner or an admin may change the managers.
list_apps	Apps with phase, URL, owner; optional namespace filter. Inline sources list file paths only.
get_app	Full spec + status for one app; inline files come back fenced (see below).
get_app_status	Lightweight poll: phase, URL, replicas, conditions.
get_app_logs	Recent pod logs (fenced, last 64 KiB) — first stop when phase is Failed.
update_app	Patch spec (source, runtime, access, …). access.managers ({users, groups}) is the manager list; only the owner or an admin may change it. secrets ({NAME: value} sets or rotates, {NAME: null} removes) changes the app’s secret values.
stop_app / start_app	Scale to zero / back up; start_app also wakes an Idle app.
restart_app	Roll the app’s pods without changing the spec; fails if the app is stopped.
get_app_metrics	Per pod: instantaneous CPU/memory usage (available:false, and no usage, when the cluster has no metrics server), plus the app container’s restarts, last termination (reason such as OOMKilled, exitCode, finishedAt), startedAt, and CPU/memory limits and requests, which come from the pod either way.
remove_app	Delete permanently (cascades). Refused unless called with confirm: true, which the agent sets only after the user agrees.

Every tool carries MCP annotations, so clients that auto-approve tools can tell them apart: describe_cluster, list_apps, get_app, get_app_status, get_app_logs and get_app_metrics are readOnlyHint; launch_app (it updates an existing app), update_app, stop_app and remove_app are destructiveHint.

Resources

Clients that support MCP resources can attach an app’s spec, status, logs or events as context without calling a tool. URIs use the server’s own scheme, nebari-apps://, so they cannot collide with another server’s, and their paths mirror the REST API (/api/v1/apps/{namespace}/{name}/…):

URI	Content (JSON)	Same as
nebari-apps://cluster	What the cluster supports: source types, namespaces, apps domain, idle timeout, image registries.	describe_cluster
nebari-apps://apps	The apps you own or manage (all apps for admins); inline sources list file paths only.	list_apps
nebari-apps://apps/{namespace}/{name}	One app’s full spec and status, with your role; inline files fenced.	get_app
nebari-apps://apps/{namespace}/{name}/status	Phase, URL, replicas, conditions, message.	get_app_status
nebari-apps://apps/{namespace}/{name}/logs{?lines,container}	Recent pod logs, fenced (last 64 KiB). lines is 1–5000 (default 200); container defaults to app, and git-clone, copy-source or proxy read the init containers or the sidecar, e.g. …/logs?lines=500&container=git-clone.	get_app_logs
nebari-apps://apps/{namespace}/{name}/events	Kubernetes events for the app and its workload, newest first, fenced as one JSON document (capped at 32 KiB).	—

resources/list returns the two fixed resources plus one nebari-apps://apps/{namespace}/{name} entry per app you own or manage, fetched for each request with your token; the other three are resource templates. An entry’s description carries only its phase and your role, never text the app’s author wrote. Secret values never appear in any resource: like the tools, they show only the names in runtime.secretKeys. A read you are not authorized for fails with the API’s 403 or 404. Subscriptions (change notifications) are not supported yet; re-read …/status to poll.

Prompts

Prompts package the common workflows. Each returns one message that tells the agent which resources to read, in order (with the matching tool for clients without resource support), instead of inlining logs or specs:

Prompt	Arguments	What it does
triage_app	namespace, name	Diagnose a failing or stuck app in the order Troubleshooting uses: status, the first False condition, events, then logs (including the init containers), then propose a fix and wait for the user before changing anything.
launch_from_manifest	path (default nebari-app.yaml)	Launch or update an app from a nebari-app.yaml manifest: map it onto launch_app, check it against nebari-apps://cluster, keep secret values out of env, poll the status resource until Running.
review_access	namespace, name	Summarize who can open the app (public, signed-in users, or restricted to groups/users) and who can manage it (owner, access.managers, admins), your role, landing-page and env exposure, and propose an update_app patch for the user to approve.

In Claude Code, prompts appear as slash commands, for example /mcp__nebari-apps__triage_app apps docs-site.

Untrusted output

Pod logs, inline source files and Kubernetes events (whose messages quote container output) are written, directly or not, by whoever launched the app, and they land in the agent’s context. Tools and resources return them inside markers with a random nonce (<untrusted-1a2b…> … </untrusted-1a2b…>) that the content cannot close early, next to a notice telling the agent to treat the content as data, never as instructions. Logs are capped to the last 64 KiB, events to the first (newest) 32 KiB and each inline file to its first 16 KiB (truncated / truncatedFiles say when).

Secrets

launch_app and update_app accept secrets, a map of environment variable names to credentials. apps-api writes them to the app’s own Kubernetes Secret and returns only the names (runtime.secretKeys), so no tool, get_app and list_apps included, ever outputs a value. Values an agent passes still appear in its own transcript, which coding agents often log: the tool descriptions steer agents to ask the user to type credentials into the UI’s Secrets section instead, and never to put them in env.

Errors

Every tool returns structured JSON. Errors carry the API’s human-readable detail, and a 401 includes a nextAction telling the agent to have the user sign in to the server again from their MCP client (in Claude Code: /mcp, then Authenticate).

Clients that use the handshake-era protocol (before 2026-07-28) keep an MCP session in the server’s memory, so mcp.replicaCount stays 1. Authentication does not depend on it.

Example agent flow

“Launch the site in ./docs-site as a public app called docs-site”

1. describe_cluster → namespaces ["apps", ...], appsDomain apps.example.ai
2. launch_app with source_type: inline, the directory’s files as inline_files, public: true, subdomain: docs-site (add landing_page: true to get a tile on the Nebari landing page; apps are hidden from it by default)
3. get_app_status until phase: Running
4. Report https://docs-site.apps.example.ai back to the user

For a Python app (“launch the pixi app in ./py-app with task start”), step 2 adds pixi_task: "start" and the files must include a pixi.toml (or pyproject.toml).

Pair it with the scaffolding skill: the skill generates apps with a nebari-app.yaml manifest, and “launch it” maps that manifest onto launch_app with no ambiguity. Agents without the skill get the same steps from the launch_from_manifest prompt.