Configuration

TOML config with sensible defaults — credentials in muninn.toml.local (gitignored)

Config Precedence

1. MUNINN_CONFIG env var → path to config file
2. Falls back to muninn.toml in working directory
3. muninn.toml.local is gitignored — used for real credentials
4. Systemd service sets MUNINN_CONFIG=muninn.toml.local

[matrix]

Key	Type	Description
`homeserver`	String	Matrix homeserver URL
`user_id`	String	Bot's Matrix user ID
`password`	String	Bot's password
`store_path`	String?	E2EE crypto store path
`sync_timeout`	u64?	Sync timeout in seconds (default: 10)
`allowed_users`	[String]?	Whitelist of user IDs (empty = allow all)

[llm]

Key	Type	Description
`api_url`	String	OpenAI-compatible API endpoint (e.g. the GPU server :30123)
`model`	String?	Model name (default: "default")
`classifier_url` / `classifier_model`	String?	Secondary endpoint for short latency-sensitive prompts (topic classifier, profile slicer). Defaults to `api_url`/`model`.
`system_prompt`	String?	Inline base prompt (legacy). Overridden by workspace files.
`workspace_path`	String?	Dir of prompt markdown (IDENTITY/SOUL/RULES/USER/TOOLS). Default `~/.config/muninn/workspace`.
`max_tokens`	u32?	Max response tokens (default: 2048)
`temperature`	f64?	Sampling temperature (Qwen: 0.6, Gemma: 1.0)
`top_p`	f64?	Nucleus sampling (default: 0.95)
`top_k`	u32?	Top-k sampling (Qwen: 20, Gemma: 64)
`presence_penalty`	f64?	Presence penalty for repetition control
`reasoning_budget`	i32?	Per-request thinking budget (llama.cpp). 0 = off.
`enable_thinking`	bool?	Thinking toggle (vLLM/Qwen `chat_template_kwargs`).
`log_tool_results`	bool?	Include tool args/results in `data/logs/` dumps (default false — privacy).
`replay_capture`	bool?	Write replayable JSONL per session to `data/logs/replay/`, last 5 per channel (default true).

LLM Model Notes

Model	temp	top_k	reasoning	Notes
Qwen3.6 27B	0.6	20	off	Thinking breaks tool calls (llama.cpp #20837). `--jinja` required.
Gemma 4 31B	1.0	64	on (budget 128)	Google recommends thinking for tool accuracy. `--jinja` required.

[memory]

Key	Default	Description
`db_path`	`./data/muninn.db`	SQLite database path
`recent_messages`	10	Messages in context window
`max_memories`	5	Memories injected per query
`extraction_enabled`	true	Background fact extraction
`consolidation_enabled`	true	Background maintenance (episodes, decay, merge)
`consolidation_interval_secs`	1800	Maintenance cycle interval (30 min)
`decay_half_life_days`	14.0	Ebbinghaus decay half-life
`decay_threshold`	0.1	Archive memories below this strength
`episode_window_minutes`	60	Time window for grouping episodes
`ollama_url`	None	Ollama API URL (enables embeddings)
`ollama_model`	nomic-embed-text	Ollama embedding model
`skill_extraction_enabled`	true	Distil successful sub-agent trajectories into skills
`skill_curator_enabled` + `skill_archive_unused_after_days` / `skill_archive_idle_after_days` / `skill_min_uses_for_quality_check` / `skill_min_success_rate`	true / 30 / 90 / 5 / 0.3	Skill janitor: archive unused/idle/low-success skills (v0.12)
`profile_enabled` + `profile_min_memories` / `profile_max_chars` / `profile_refresh_after_hours`	true / 10 / 1500 / 24	Per-user dialectic profile (Honcho-lite, v0.12)
`profile_slicer_enabled` + `profile_slice_max_chars`	false / 400	Query-aware profile slicing on the hot path (extra LLM call/turn)
`dream_enabled` + `dream_interval_hours` / `dream_max_messages`	false / 24 / 120	Opt-in nightly reflective consolidation pass (v24)
`peer_db_path`	None	Read-only peer DB for shared-memory lookup (TUI surface)

[tools]

Key	Type	Description
`searxng_url`	String?	SearXNG URL for web search tool
`search_max_results`	usize?	Max search results (default: 5)
`shell_enabled`	bool?	Enable shell tool (default: false — opt-in only)
`shell_timeout_secs`	u64?	Shell command timeout (default: 30)
`shell_ssh_remote_allowlist`	[String]?	Hosts the shell tool may SSH to (empty = all outbound denied)
`shell_sandbox_hosts`	[String]?	SSH targets that skip safety gates (sandbox)
`calculate_enabled`	bool?	Enable calculate tool (default: true)
`calculate_timeout_secs`	u64?	Calculate timeout (default: 10)
`ha_url`	String?	Home Assistant URL (enables weather, home, camera, music tools)
`ha_token`	String?	HA long-lived access token
`ha_allowed_users`	[String]?	Matrix IDs allowed HA tools (AND-ed with role gate; empty = all)
`music_assistant_config_entry_id`	String?	HA Music Assistant entry id — enables library search + "play some music"
`music_default_player`	String?	Fallback media_player when no player named and room has no audio_out
`comfy_url`	String?	ComfyUI URL for image generation (enables imagine tool)
`voice_server_url`	String?	Kai voice server URL (enables speak tool)
`stt_url` / `stt_token` / `stt_language`	String?	STT endpoint (enables transcribe tool)
`sandbox_path`	String?	Dev-tools sandbox dir (default: /tmp/muninn/sandbox)
`[[tools.mcp_servers]]`	table[]	MCP servers — `kind = "stdio"` (command/args/env) or `"http"` (url/headers). Admin-gated.

Page fetch is no longer a native tool — the removed web_fetch/firecrawl_url are replaced by the muninn-fetch HTTP MCP server.

[tools.email]

Enables the email tool (admin-only). Master enabled toggle plus one [[tools.email.accounts]] block per mailbox. Defaults suit Mailcow: IMAP 993 implicit-TLS, SMTP 587 STARTTLS. Credentials live in muninn.toml.local.

Key	Type	Description
`enabled`	bool?	Master toggle. Absent/false → tool not registered.
`[[tools.email.accounts]]`	table[]	One per mailbox. Fields below.
↳ `id` / `name` / `default` / `enabled`	String / String? / bool? / bool?	Selector the LLM names; label; default account; per-account toggle
↳ `imap_host` / `imap_port` / `imap_security`	String / u16? / String?	IMAP host; port (993); `ssl`\|`starttls`\|`none`
↳ `smtp_host` / `smtp_port` / `smtp_security`	String? / u16? / String?	Defaults to imap_host; port (587); `starttls`\|`ssl`\|`none`
↳ `user` / `password`	String / String	Login for BOTH IMAP and SMTP
↳ `from_address` / `from_name` / `archive_folder` / `trash_folder`	String?	From (defaults to user); display name; Archive; Trash

[tools.calendar]

Enables the calendar tool + reminder worker (admin-only). One shared calendar; events stored locally and best-effort mirrored to SOGo CalDAV. Recurring events supported.

Key	Type	Description
`enabled`	bool?	Master toggle (gates the tool AND the reminder worker).
`default_reminders`	[String]?	Lead times when the user doesn't specify (default `["1 day", "1 hour"]`). Each is "<N> minute\|hour\|day\|week".
`[[tools.calendar.accounts]]`	table[]	One per calendar. Fields below.
↳ `id` / `name` / `default` / `enabled`	String / String? / bool? / bool?	Selector; label; default; per-account toggle
↳ `caldav_url` / `user` / `password`	String	CalDAV collection URL (e.g. `…/SOGo/dav/<user>/Calendar/personal/`) + basic-auth creds

[tools.contacts]

Enables the contacts tool (admin-only) — CardDAV name→email/phone lookup over a single address book (e.g. SOGo's Personal Address Book; reuses the mailbox creds).

Key	Type	Description
`enabled`	bool?	Master toggle. Absent/false → tool not registered.
`carddav_url`	String	CardDAV collection URL, e.g. `…/SOGo/dav/<user>/Contacts/personal/`
`user` / `password`	String	Basic-auth credentials (muninn.toml.local)

[tools.youtube]

Enables the youtube tool (admin/adult) — transcript→summary via yt-dlp. Requires the yt-dlp binary on the host.

Key	Type	Description
`enabled`	bool?	Master toggle. Absent/false → tool not registered.
`binary`	String?	Path to `yt-dlp` (default: resolved via PATH).
`default_lang`	String?	Preferred caption language (default `en`).

[tools.research]

Enables the research tool (admin/adult) — deep plan→search→fetch→synthesize into a cited brief, over the top-level [tools] searxng_url. Slow → run backgrounded (!bg).

Key	Type	Description
`enabled`	bool?	Master toggle. Absent/false → tool not registered.
`max_subquestions`	usize?	Sub-questions the planner expands into (default 3, max 6).
`results_per_search`	usize?	Results fetched per sub-question (default 4, max 10).
`max_page_chars`	usize?	Per-page text cap fed to synthesis (default 4000).
`fetch_pages`	bool?	Fetch + read result pages (default true); false = snippets only.

[docs]

Key	Default	Description
`port`	7790	Port for the embedded docs web server