Settings Reference

The settings page is split into virtual tabs so setup and advanced features stay separate.

Models

Configure provider, API key, model id, temperature, max tokens, timeouts, retries, and model-network proxy settings.

Providers:

• OpenRouter
• Ollama
• OpenAI-compatible

Agent

Configure standing instructions, output style, compaction, memory, skills folders, and runtime resource behavior.

• Permission mode — Safe honors approval gates; YOLO auto-approves mutating tools for the session. Plan mode is entered via /plan in chat.
• Temperature — Sampling randomness (0–2).
• Max response tokens — Per model request. 0 lets the provider decide.
• Request timeout — How long to wait for the provider to start responding.
• Network retries — Automatic retries on rate limits and transient errors.
• System prompt — Sent at the start of every conversation.
• Standing instructions — AGENTS.md (or CLAUDE.md / GEMINI.md) loaded from the vault root every turn.
• Context window — Auto-compaction settings: summarize older turns automatically as the context window fills.
• Tool budget — Drop optional tools when registered tool schemas exceed a threshold percent of the context window.
• Subagents — Subagent timeout (auto-abort a child after N seconds; 0 disables, max 86400).

Context-window resolution

OpenRouter models use the live /models catalog for their context window. For the OpenAI-compatible provider, the plugin tries the OpenRouter catalog for the same model slug (exact, or an unambiguous suffix match); otherwise the window is treated as unknown. Unknown windows keep every tool (the tool budget never drops) and disable auto-compaction rather than guessing 128k and silently dropping optional tools. If your gateway exposes a window different from OpenRouter's, set the provider's Context window (tokens) setting (0 = auto-detect).

Approval

Set the default mode, mutating-tool policy, per-tool overrides, working directories, and ignore patterns.

• Before mutating tools — Global allow/ask/deny for tools that change the vault.
• Working directories — Grant folders as a working set. Inside granted folders reads and writes auto-run; outside asks first.
• Per-tool overrides — Allow, ask, or deny for individual tools.

Web

Enable web access and configure Tavily, Brave, or SearXNG credentials and endpoints, plus an optional destination allowlist.

• Search provider — Tavily, Brave, or SearXNG.
• Max results — Default search results to return (1–10).
• Fetch character limit — Cap on characters of fetched page text returned to the model.
• Fetch allowlist (settings.web.allowedHosts, comma-separated host suffixes, e.g. example.com, *.wikipedia.org, *) — limits fetch_url destinations with label-boundary-aware matching; empty allows all public hosts. Deny wins over SSRF and on every redirect hop. Web/MCP/vault tool outputs are wrapped as untrusted data ([BEGIN_UNTRUSTED_TOOL_OUTPUT ...] / [END_UNTRUSTED_TOOL_OUTPUT], inner markers escaped) so the model does not follow injected instructions inside tool results.

MCP

Enable remote MCP, generate agent plugin packages for HTTPS Streamable HTTP servers, choose auth, test discovery, and set per-server approval policy.

MCP servers now live inside agent plugins (.agentic-plugins/ by default) — the plugins folder is the single source of truth. The MCP tab's Add MCP server form writes a real plugin package (plugin.json + mcp.json); endpoint and headers come from the package, while authentication, approval, and enable state stay client-owned.

Supported auth modes:

• none
• bearer token
• custom static header
• MCP OAuth

Observability

Enable trace export, choose Langfuse or generic OTLP, set endpoint and auth, choose payload mode, and configure observability-specific proxy settings.

Payload modes:

Mode	Sends
Metadata only	Turn, model, tool, approval timing, token and cost totals, and errors.
Redacted text previews	Metadata plus short masked prompt and answer previews.
Full prompt/output content	Full prompt and answer text. Use deliberately.

Notifications

Configure cost alerts, spend caps, and related usage notifications.

• Cost alert — Notify once when session cost crosses a USD amount.
• Cost cap — Hard cap: block new turns once session cost reaches this USD amount.

Resources

Inspect runtime resources such as agent plugins, MCP tools, artifacts, retrieval state, and diagnostics surfaced by the plugin.

Also includes:

• Agent plugins — Vault folder scanned for agent plugin packages (plugin.json + skills/ + mcp.json), with per-plugin enable toggles and open-folder shortcuts. See Agent Plugins.
• Subagents — Subagent timeout.

Semantic retrieval

Opt-in semantic index configuration. Uses the same provider secrets as the Models tab.

• Embedding provider — OpenRouter, Ollama, or OpenAI-compatible.
• Embedding model — Model id for the chosen provider.
• Vector dimensions — Expected embedding vector size.
• Language coverage — Multilingual, monolingual, or unknown.
• Batch size — Maximum notes per embedding request.
• Max indexed characters per note — Upper bound sent to the embedding provider.