# Config settings (https://docs.neuralisapp.com/agent-core/config-settings)

agent-core declares **191** platform settings in its manifest's
`configSettings[]` array. They are edited in the Admin widget's **Config** tab,
which lists every key with its declaring package, its bounds and where its
current value comes from (default, environment fallback, or an admin override);
filter the list by `@neuralis/agent-core` to see exactly these. Editing needs
`platform.config`, a platform-tier feature with no default role grant. How the
store, its validation and hot reload work is on
[configuration](https://docs.neuralisapp.com/enterprise/configuration.md); the declaration contract is
[package-system config settings](https://docs.neuralisapp.com/package-system/config-settings.md).

Most edits apply on the next read, with no restart; a key whose description
says otherwise applies on restart. **Limits default generously** — the defaults
are chosen so ordinary work never meets them, and an administrator narrows them
deliberately. Secrets are never settings: provider keys, OAuth app pairs and
channel tokens live in [credentials](https://docs.neuralisapp.com/agent-core/credentials.md).

The groups below follow the Config tab's own categories; the key names are the
ones the tab shows.

## Runtime (75 keys) [#runtime-75-keys]

The agent loop and the conversation around it.

- **Agent stream** — `maxAgentSteps` (2000), `defaultModel`,
  `defaultTemperature`, `maxAgentsPerProject` (100), `resumeMaxAttempts` (chat
  auto-resume after a restart), `abortJoinGraceMs`, `perfSlowToolMs` (the
  `perf.slow` threshold for one tool call, 30 s), `shutdownDrainTimeoutMs`.
- **Tool results** — `toolResultContentCapChars` (50 000),
  `toolResultStructuredCapChars` (20 000), `toolResultMetaCapChars` (256 000):
  the cap every tool result meets before it reaches the model.
- **Shell** — `shellDefaultTimeoutMs` (10 min), `shellMaxOutputBytes`,
  `shellKillGraceMs`, the live-output trio `shellLiveFlushMs` /
  `shellLiveFrameMaxBytes` / `shellLiveStreamMaxBytes`, background shells
  (`shellBackgroundRingBytes`, `shellBackgroundNotify`, `shellReportTailChars`,
  `backgroundWakeClientGraceMs`, `shellExitedGcTtlMinutes`), and
  `shellScratchTtlHours`.
- **Subagents** — `delegateMaxDepth` (3), `delegateDefaultMaxSteps` (1000),
  `delegateDefaultTimeoutMs` (2 h), `delegateAnswerCapChars`, and the background
  trio `delegateBackgroundMaxConcurrent` (8), `delegateBackgroundNotify`,
  `delegateMaxResumeAttempts`.
- **Conversation inbox** — `conversationInboxMaxChars`,
  `conversationInboxMaxPending`, `conversationInboxMaxHops`,
  `conversationInboxMinIntervalMs`, `conversationInboxImmediatePerTurnMax`,
  `conversationEventGateTtlMs`.
- **Context and prompt** — `identityFileMaxChars` / `identityFileMaxCount`,
  `packageFileMaxChars`, `runtimeStackAreasMaxChars`,
  `sessionArchiveMaxEntries` (archived goal cycles), attachment read caps, and
  the composer paste caps `composerPasteMaxImageBytes` (25 MB) /
  `composerPasteMaxTextChars` (500 000).
- **Media** — `mediaRequestBudgetBytes`, `videoSampleMaxFrames`, and the
  ffmpeg/ffprobe timeouts.
- **Skill export** — `pluginExportMaxFiles`, `pluginExportMaxBytes`,
  `pluginExportMaxEntries`.
- **Devices** — `hostBrokerEnabled` (the kill-switch: on is necessary, never
  sufficient; off disables every device at once), probe and filesystem
  timeouts, link ping/stall/stream bounds, `deviceOfflineGraceMs` (15 s),
  `deviceLivenessTtlMs`, `deviceEventGateTtlMs`, `deviceWatchLingerMs`
  (120 s: how long a device keeps a folder watch after its link drops, so the
  redial resumes it without a full re-sync), and the background-shell trio
  `hostShellBackgroundLifetimeMs` / `hostShellPollMs` /
  `hostShellStatusWaitMs`.
- **Terminal** — `terminalMaxSessions` (per user, project and agent; `429`
  past it), `terminalScrollbackLines`, `terminalExecTimeoutMs`,
  `terminalMaxOutputBytes`, `terminalIdleTimeoutMinutes`.
- **Lifecycle hooks** — the timeouts for principal revocation, credential
  change, package health and project deprovision hooks.

## Workflow (11 keys) [#workflow-11-keys]

`workflowMaxConcurrentRuns` (16, platform-wide), `workflowMaxRuntimeMinutes`
(the 24 h ceiling) and `workflowDefaultRuntimeMinutes` (4 h),
`workflowMaxRunRetries`, `workflowRetryBaseDelayMs`,
`workflowMaxResumeAttempts` (5; 0 turns resuming off),
`workflowBusyConversationDeferMs`, `workflowSchedulerTickMs`,
`workflowMissedFireGraceMs` (15 min catch-up window), and the in-tool wait
`workflowWaitCapSeconds` (120; 0 turns waiting off) / `workflowWaitPollMs`. See
[workflows](https://docs.neuralisapp.com/agent-core/workflows.md).

## Summarizer (8 keys) [#summarizer-8-keys]

`summarizerAutoThresholdPercent` (80), `summarizerPreserveTurns`,
`summarizerManualPreserveTurns`, `summarizerManualMinIntervalMs`,
`summarizerMaxTokens`, `summarizerRequestMaxChars`, and the per-request and
total timeouts.

## Channels (3 keys) [#channels-3-keys]

`channelFinalTextCapChars`, `channelInboundDebounceMs` (2 s), and
`channelInboundFiresPerMinute` (30 per connection; 0 disables). See
[channels](https://docs.neuralisapp.com/agent-core/channels.md).

## Web (31 keys) [#web-31-keys]

- **Search** — `websearchProviderOrder` (`tavily,brave,native,searxng`),
  `websearchProviderNative`, `websearchSearxngUrl` / `websearchSearxngLanguage`,
  provider timeout, failure-disable and cooldown, `websearchNativeMaxTokens`,
  fusion and freshness ranking (`websearchRrfK`,
  `websearchFreshnessHalfLifeDays`), and the research-mode budget
  (`websearchResearchDepth`, `websearchResearchPaidVariantCap`, the content
  fetch count/timeout/size, `websearchResearchByteBudget`).
- **Fetch** — `webFetchDefaultTimeoutMs`, `webFetchDefaultMaxChars`,
  `webFetchExcerptMaxChars`, `webFetchPdfMaxBytes`,
  `webFetchJsOnlyTextFloorChars`, `webFetchMaxResponseBytes`,
  `webFetchMaxRedirects`, `webFetchNetworkRetries`, and the outbound default
  for packages, `webFetchTimeoutMs`.
- **Extraction** — `webExtractTimeoutMs`, `webExtractWorkers`,
  `webExtractWorkerMaxHeapMb`, `webExtractWorkerIdleMs`.
- **Egress** — `webToolsAllowedHosts`, `webToolsUserAgent`.

## MCP (18 keys) [#mcp-18-keys]

Sessions on the hosted MCP endpoint (`mcpMaxSessionsPerUser`,
`mcpMaxSessionsTotal`, idle timeout, max lifetime, sweep interval, SSE
keep-alive, `mcpAllowedOrigins`), the outbound client's reconnect back-off and
stdio RPC timeout, and the managed sidecar (`mcpSidecarImage`, port, idle
minutes, memory, CPUs, process limit, health timeout).

## Providers (9 keys) [#providers-9-keys]

`localLLMs` (the custom endpoints — see
[providers and models](https://docs.neuralisapp.com/agent-core/providers-and-models.md#adding-a-custom-endpoint)),
the key and model caches, custom-endpoint discovery and stream timeouts, and
the transient-retry budget (`transientRetryMax`, base and cap).

## Advanced (34 keys) [#advanced-34-keys]

Per-provider request timeouts and base URLs, `geminiStoreRequests`,
`anthropicDefaultThinkingBudget`, `providerTopP` / `providerTopK`,
`billingMaxCompletion`, the Codex WebSocket switches
(`codexWebsocketEnabled`, idle and frame timeouts, `codexWebsocketMaxSockets`),
trusted-package rate limits and timeouts, `projectServicesCacheSize`,
`credentialExpiryWarningHours`, `logsTailMaxLines`, and the package log keys
(`packageLogLevel`, `packageLogLevelOverrides`, size, file count and age) —
those apply live, on the next line written.

## Debug (2 keys) [#debug-2-keys]

`streamTimingEnabled` (the per-step timing table) and `anthropicLog`.
