LobeHub
LobeHub

Feature Flags

In addition to basic environment variable configuration, LobeHub also offers feature flags to control whether a feature is enabled globally, or to enable or disable features for specific user groups or environments as needed.

Check Your Deployment Version

Flag names and defaults can change between releases. Check the schema at your deployed version, not only the latest main branch. Feature status labels describe availability; they are not flag names. This guide does not provide an unverified flag for interactive Agent sharing.

Feature Flags Environment Variable FEATURE_FLAGS

  • Type: Optional
  • Description: Used to control LobeHub's feature functionalities. Supports multiple feature flags, using + to add a feature and - to disable a feature. Separate multiple feature flags with a comma , and enclose the entire value in quotes " to avoid parsing errors.
  • Default: -
  • Example: "-welcome_suggest"

FEATURE_FLAGS controls the Boolean flags listed here. It does not replace service settings, permissions, credentials, or deployment requirements. An enabled flag alone does not guarantee that a feature is available.

You can achieve various feature combinations using the above configuration syntax. All feature flags are Boolean values, enabled with + and disabled with -.

Attention: Unlike the OPENAI_MODEL_LIST variable, the FEATURE_FLAGS variable does not support the all keyword. You need to manually control all feature flags (otherwise, they will adopt their default values).

Configuration ItemDescriptionDefault Value
check_updatesAllows checking for updates.Enabled
provider_settingsControls model provider settings display.Enabled
openai_api_keyAllows users to customize the OpenAI API Key.Enabled
openai_proxy_urlAllows users to customize the OpenAI proxy URL.Enabled
api_key_manageControls access to API key management page (/profile/apikey).Disabled
agent_shareGates new or re-enabled interactive Agent shares. Requires deployment support; see the boundary below.Disabled
agent_llm_relayRuns LLM calls for model providers only the user's device can reach (local Ollama / LM Studio, private-network endpoints) on the client that started the run, while the agent loop stays on the server. Requires an open LobeHub client for such runs.Disabled
agent_deferred_initReturns a plain chat send as soon as its messages are saved and runs tool discovery and context assembly in the first agent step, so sending feels instant. Requires the queue agent runtime.Disabled
agent_deferred_hetero_dispatchReturns a plain chat send to a heterogeneous agent (Claude Code, Codex, …) on an online device before the device accepts the run; the dispatch runs after the response, and a failure is shown on the assistant message.Disabled
dashboardDashboards: the home and project dashboard pages, the widget Portal, in-chat widget previews and the builtin-dashboard agent tool that builds monitoring widgets. When off, these are hidden and the widget / dashboard APIs return FORBIDDEN; widgets already published keep refreshing on their schedule.Disabled
mcp_appsMCP Apps: a tool result whose MCP server declares an interactive ui:// view renders that view (the server's HTML, in a sandboxed iframe) in the conversation. When off, such results render as plain tool results and the server does not announce MCP Apps support to MCP servers.Disabled
webhookInbound webhooks wake Agents. MCP Events: an Agent subscribes to remote connector events and runs in the original Topic; subscribing follows the user's tool approval settings. Agent webhooks: URLs under Agent Profile → Webhook that any outside system calls to run the Agent. Each event or call uses credits. Requires a public APP_URL and secret encryption; MCP Events also needs an authenticated MCP server offering webhook events. When off, the entries are hidden and deliveries are still recorded but run nothing.Disabled
edit_agentAllows users to edit assistants.Enabled
ai_imageControls AI image generation feature and page (/image).Enabled
speech_to_textEnables speech-to-text functionality.Enabled
token_counterReserved for token counter display.Enabled
welcome_suggestDisplays welcome suggestions.Enabled
changelogControls changelog modal/page display.Enabled
marketEnables the assistant market functionality.Enabled
knowledge_baseEnables the knowledge base functionality.Enabled
rag_evalControls RAG evaluation feature (/repos/[id]/evals).Disabled
cloud_promotionControls cloud service promotion link display in user menu.Disabled
commercial_hide_githubHides GitHub-related links in settings footer (requires commercial license).Disabled
commercial_hide_docsHides documentation and help menu including changelog, docs, and feedback (requires commercial license).Disabled

Interactive Agent Sharing: Deployment and Rollout Boundaries

The following behavior is verified against commit 659db0780b263f023019337a48f23e17e77e260d. Other versions can differ.

The agent_share schema sets the default to false. This flag controls rollout for creators, not visitor access to existing shares.

The server gate applies two separate checks:

  1. Deployment support: ENABLE_BUSINESS_FEATURES must be true. Without it, both creation and visitor access fail with FORBIDDEN before the server evaluates the rollout flag.
  2. Creator rollout: New or re-enabled shares also require enableAgentShare === true for the creator. A false or undefined value fails closed. Array values in runtime configuration identify permitted creators, not visitors.

In the verified open-source code, ENABLE_BUSINESS_FEATURES is a code constant set to false, not an environment variable. FEATURE_FLAGS="+agent_share" cannot bypass this deployment gate. This flag does not enable interactive Agent sharing on a standard self-hosted deployment.

The business constant defines this boundary. This guide does not prescribe a code change to bypass it.

On a supported deployment, removal from the creator rollout does not itself revoke existing shares. Visitors do not need the rollout flag to access or chat through an existing share. Owners can still manage, preview, or revoke existing shares. These operations remain subject to their existing authorization checks.

The share router checks the creator gate for enableShare and updateVisibility to link. A change to private does not require the rollout flag and blocks visitor access. Thus, disabling agent_share is not a substitute for disabling an individual share.

You can always check the featureFlags to get the latest list of feature flags.

Agent Webhooks: Deployment Requirements

Agent webhooks and MCP Events share the webhook flag, off by default. Before enabling it for Agent webhooks:

  1. Set APP_URL to the public address callers will use; webhook URLs are built from it.
  2. Configure KEY_VAULTS_SECRET to encrypt webhook keys, and preserve it across deployment updates.
  3. Optionally configure QStash so wake-ups and recovery of pending calls run on a schedule. Without it, wake-ups run in the application process and recovery depends on application requests.
  4. Set FEATURE_FLAGS="+webhook" and restart.

Open an Agent, go to Agent Profile → Webhook, create a webhook and use Send test: the call should show as run under Recent calls, with a link to the topic the Agent replied in.

Turning the flag off hides the Webhook and Automations tabs and stops webhook-triggered runs. Calls are still authenticated, stored and acknowledged, but skipped without using credits; webhooks and their URLs are kept and resume working when the flag is turned back on.

MCP Events: Deployment Requirements

MCP Events needs more than Agent webhooks. Before enabling webhook for it:

  1. Set APP_URL to a public HTTPS address the MCP server can reach. Localhost, private hostnames and IP-literal addresses are unsupported.
  2. Configure KEY_VAULTS_SECRET to encrypt webhook signing secrets. Generate it using the server environment variable guide, and preserve it across deployment updates.
  3. Configure QStash for scheduled renewal, unsubscribe retries and pending-event recovery. The Docker launcher registers a sweep every five minutes; other deployments need their own schedule. Without QStash, maintenance depends on application requests and has no guaranteed timing.
  4. Set FEATURE_FLAGS="+webhook" and restart. Connect an authenticated remote HTTP MCP server offering SEP-3415 webhook events, then refresh its connector tools.

Ask the Agent to subscribe in a conversation and approve it. Open Settings → Automations and confirm the subscription is active. Trigger an upstream event and confirm an event message and Agent reply appear in the original Topic. No-expiry subscriptions have no scheduled health check; use Check in Automations to restore interrupted delivery.

Turning the flag off hides the tools and Automations entry, and stops event-triggered Agent runs and subscription renewals for affected users. Incoming events are still verified and stored, but are skipped without using credits. Subscription records remain; finite subscriptions may expire while the flag is off.

Re-enabling the flag restores event runs for valid subscriptions and lets scheduled renewals attempt to recover expired subscriptions from the saved cursor. You can delete unwanted subscriptions before or after disabling the flag.