32+ Slash Commands

Type / in the WebUI/TUI to trigger the command palette, or press Ctrl+K in the web dashboard for the same commands. You can also type a command directly like /model claude-sonnet-4-6.

Identity files

These Markdown files live in ~/.curie-agent/. Created by /init, read at every session start. Edit them directly in any text editor — they're plain Markdown.

File	Purpose	When updated
`AGENTS.md`	Agent registry and role definitions — injected into context every session	Manually
`SOUL.md`	Personality, voice, operating principles, behavioral constraints	Manually
`USER.md`	Your profile: name, timezone, goals, preferences, expertise	Manually + heartbeat
`MEMORY.md`	Long-term knowledge: facts, heuristics, project history, lessons learned	Weekly / monthly heartbeat
`HEARTBEAT.md`	Proactive routine definitions for all heartbeat cycles	Via `/heartbeat` commands
`TOOLS.md`	Local service/device registry — SSH hosts, IoT endpoints, custom paths	Manually
`WIKI.md`	Wiki schema and Ingest/Query/Lint procedures — scaffolded at `wiki init`, injected on demand	Manually / wiki init

Built-in skills

Skills are activated by prompt or slash command. Find them in ~/.curie-agent/skills/ after running /init. Use /skill to list all available skills.

`/deep-research`

Multi-source structured research with citation tracking and synthesis.

Four-phase methodology

Orientation — 2–4 broad searches to map the landscape
Targeted Deep — 5–15 narrow searches on key sub-questions
Verification — 2–5 counter-searches to challenge findings
Gap-fill — additional searches as needed

Output formats: research summary, comparison table, deep report, annotated source list, fact-check.

`/planning`

Universal planning tool — gathers context before committing to a plan.

Seven planning modes

Roadmap — phased milestones for big goals
Timeline — work backwards from a deadline
Goal Decomposition — break overwhelm into actions
Project — multi-person deliverables with risks
Periodic — weekly/monthly time-blocking
Event — who/what/when/where/how
Habit — sustainable systems and routines

`wiki ingest / query / lint`

Persistent, compounding knowledge base. Ingest external sources, query with citations, lint for health. Storage: ~/.curie-agent/wiki/

Three workflows

Ingest — reads source → writes summary page → updates entity/concept pages → updates index → appends log
Query — reads index → searches pages → synthesizes answer with [[citations]] → files novel discoveries back
Lint — deterministic checks (orphans, broken links, missing-from-index, stale) → semantic contradiction check → fixes

Run via CLI verbs (curie-agent wiki <cmd>) or natural language in the TUI.

Settings reference

Stored in ~/.curie-settings.json. Configure via slash commands or edit directly. CLI flags override settings; settings override hardcoded defaults.

Key	Default	Description
Core
`model`	`claude-sonnet-4-6`	Active model. Set via `/model`
`effort`	`auto`	Reasoning effort: low, medium, high, max, auto. Set via `/effort`
`mode`	`auto`	Approval mode: plan, edit, auto, yolo. Set via `/mode`
`theme`	`tokyo-night`	Color theme (8 options). Set via `/theme`
Safety
`SAFETY_PATH_GUARD`	`on`	Refuse file ops outside project CWD + allowlist
`SAFETY_COMMAND_GUARD`	`on`	Hard-deny lethal shell patterns; force-ask risky ones
`SAFETY_SNAPSHOTS`	`on`	Auto git stash snapshots before yolo-mode edits
Heartbeat
`HEARTBEAT_DAILY`	`07:15`	Morning briefing time (HH:MM). Set via `/heartbeat daily`
`HEARTBEAT_DREAMING`	`23:01`	Nightly synthesis time. Set via `/heartbeat dreaming`
`HEARTBEAT_WEEKLY`	`Fri 21:00`	Weekly review time. Set via `/heartbeat weekly`
Channels
`TELEGRAM_BOT_TOKEN`	—	Telegram bot token. Set via `/channels set-bot-token`
`TELEGRAM_USER_ID`	—	Your Telegram user ID. Set via `/channels set-user-id`
Tool limits
`TOOLS_PER_CALL`	`10`	Max tool calls per turn. Set via `/tools`
`WEBSEARCH_PER_CALL`	`5`	Max web searches per turn. Set via `/websearch`
Wiki
`wiki.path`	`''`	Wiki storage path. Empty = `~/.curie-agent/wiki/`
`wiki.autoLint`	`off`	Auto-lint wiki during heartbeat cycles. Set to `on` to enable.

Wiki CLI verbs

The wiki is a top-level CLI subcommand. Run these from your terminal in addition to using natural language in the TUI. Initialise once with wiki init, then ingest sources and query as your knowledge base grows.

Command	Description
`curie-agent wiki init`	Scaffold wiki structure in `~/.curie-agent/wiki/` — creates directories, `WIKI.md` schema, empty index and log
`curie-agent wiki ingest <path\|url>`	Add a source — agent reads it, writes summary and entity/concept pages, updates the index, appends log
`curie-agent wiki query "<question>"`	Answer a question from the wiki with cited pages. Files novel discoveries back automatically.
`curie-agent wiki lint`	Health check: orphan pages, broken wikilinks, missing-from-index entries, stale claims. Agent fixes what it finds.
`curie-agent wiki graph`	Emit the backlink graph as JSON — nodes are pages, edges are `[[wikilinks]]`

`/status`

Display current session status — version, active model, provider, approval mode, CWD, token counts, tool limits, and pricing.

Usage

/status

Passing no arguments shows all current values. The output is also shown live in the status line at the bottom of the TUI.

`/help`

List all slash commands organized by category with their usage syntax and descriptions.

Usage

/help

`/model`

Switch the active AI model. Supports aliases for quick switching, custom pricing, and context window adjustment.

Usage

/model claude-sonnet-4-6

alias Short names: opus, sonnet, haiku, gpt4o, gpt4turbo, o1, o3-mini

pricing Set custom per-million pricing: /model pricing 0.5;2.0 (in;out)

windows Set max context window in tokens: /model windows 1000000

Default model is claude-sonnet-4-6. Changing the model also resets token counters and context window.

`/provider`

Switch the AI provider. Each provider uses its own API key and base URL from settings.

Usage

anthropic Uses ANTHROPIC_API_KEY (+ optional ANTHROPIC_URL for custom endpoints)

openai Uses OPENAI_API_KEY (+ optional OPENAI_URL)

google Uses GOOGLE_API_KEY (+ optional GOOGLE_URL)

local Uses MODEL_URL + MODEL_API_KEY for any OpenAI-compatible local server

ollama Uses MODEL_URL + MODEL_API_KEY (Ollama's OpenAI-compatible API)

openrouter Uses OPENROUTER_API_KEY + OPENROUTER_URL — routes through Anthropic's API adapter

`/mode`

Set the approval mode — controls which tool calls require user confirmation and which run automatically.

Usage

/mode <plan|edit|auto|yolo>

plan Read-only tools free. Writes only allowed under plans/ directory. All other mutations denied.

edit Edit and Write auto-approved. Shell commands and network calls ask for approval.

auto Non-destructive edits auto-proceed. Each tool call is first checked by the LLM for safety. Network calls ask.

yolo No approvals at all. Everything runs automatically. Use with caution.

`/theme`

Change the TUI color theme. Changes persist across restarts.

Usage

/theme <name>

themes tokyo-night (default), nord, dracula, solarized, gruvbox, black, white, grey

`/effort`

Set the extended thinking token budget. Controls how many tokens the model can use for internal reasoning before answering.

Usage

/effort <low|medium|high|max|auto>

low 2,000 reasoning tokens

medium 6,000 reasoning tokens

high 16,000 reasoning tokens

max 32,000 reasoning tokens

auto Let the model decide dynamically (default)

Extended thinking only works with providers that support it (Anthropic, some OpenRouter models). Higher budgets use more context window and cost more.

`/tools`

View or set the maximum number of tool calls the model can make in a single turn.

Usage

/tools [tools_per_call [websearch_per_call]]

no args Show current limits (default: 10 tools, 5 web searches per turn)

/tools 15 Set tools per call to 15 (web search limit unchanged)

/tools 15 8 Set both limits at once

`/websearch`

Set the limit for combined WebSearch + WebFetch calls per turn. Uses Brave Search API.

Usage

/websearch <count>

no args Show current limit (default: 5)

/websearch 3 Set to 3 web searches per turn

`/context`

View context window usage, list conversation messages, or compact (summarize) the conversation to free up space. Includes autocompaction and pricing tier warnings.

Usage

/context [auto|messages|compact [detailed|brief]]

no args Show context window visual bar with token count and percentage

/context messages Print the full conversation history with role labels (truncated at 8000 chars)

/context compact Summarize the conversation to free context (default: detailed summary)

/context auto Show or configure autocompaction settings and thresholds

Autocompaction

/context auto on/off — enable or disable automatic compaction when context fills up

/context auto threshold <N%> — compaction fires when context reaches N% (default: 75%)

/context auto warn <N%> — show warning at N% fill (default: 60%)

/context auto pricing on/off — toggle pricing tier warnings when cost increases after compaction

`/stats`

Switch to the Stats tab to view daily usage, session count, streaks, and token cost trends.

Usage

/stats

This command switches the TUI to the Stats tab. The tab also shows cost estimates in USD based on current model pricing.

`/memory`

Manage the agent's persistent memory files stored in ~/.curie-agent/memory/.

Usage

/memory [status|add <text>]

/memory status Show file sizes for all memory files (SOUL.md, USER.md, AGENTS.md, MEMORY.md, etc.)

/memory add Capture a memory entry. The agent will organize it into the appropriate memory files on the next turn.

`/todo`

Manage tasks with a unified tasks.json format (with legacy todo.json support). Supports manual, auto-execution, and notification modes.

Usage

/todo list Show active and done tasks with mode, priority, and schedule info

/todo add <title> Add a manual task (default: project scope)

/todo auto add /todo auto add "build at 3pm" — the agent will execute this task automatically at the scheduled time

/todo notify add /todo notify add "remind about X" — notification-only reminder with natural language time parsing

/todo complete|cancel|start|remove Change task status by ID. Requires the short task ID shown in /todo list.

Scopes

/todo personal add ... — stored in ~/.curie-agent/tasks.json

/todo project add ... — stored in the current working directory's tasks.json

`/debug`

Toggle debug logging. Shows verbose output including raw API payloads and internal events.

Usage

/debug [on|off]

Toggling with no argument flips the current state. Debug logs go to stdout and are useful for troubleshooting provider connectivity issues.

`/statusline`

Toggle the status line at the bottom of the TUI (above the input field).

Usage

/statusline [on|off]

The status line shows model, effort level, approval mode, input/output tokens, and estimated cost. Default is visible.

`/mcp`

Manage MCP (Model Context Protocol) servers. Connect external tool servers and their tools become available to the agent.

Usage

/mcp <list|add|remove|reload>

/mcp list Show all configured MCP servers with connection status (● connected, ○ disconnected, [not loaded])

/mcp add /mcp add <id> <stdio|sse|streamable-http> [--env key=value] [command] [args...]

/mcp remove /mcp remove <id>

/mcp reload Disconnect and reconnect all MCP servers

Transport types

stdio — local process (e.g. /mcp add filesystem stdio npx -y @modelcontextprotocol/server-filesystem /workspace)

sse — Server-Sent Events over HTTP (e.g. /mcp add my-api sse https://api.example.com/mcp)

streamable-http — Streamable HTTP transport

MCP servers configured in MCP_SERVERS in settings.json. Connection persists across restarts.

`/skill`

List or view Claude-Code-compatible SKILL.md files. Skills are instruction files that shape the agent's behavior for specific tasks or domains.

Usage

/skill [name]

/skill List all discovered skills with their source (global or project) and description

/skill <name> Show full details of a specific skill, including its complete instruction file

Discovery paths

~/.curie-agent/skills/ — global skills (available in all sessions)

.curie-agent/skills/ — project-level skills (available when working in that directory)

Directory format: skill-name/SKILL.md or flat: skill-name-SKILL.md

`/task`

Schedule an agent task that the LLM will automatically execute at a given time. Uses natural language time parsing and the unified TaskManager.

Usage

/task <create|list|delete>

/task create /task create "at 7:55 make a report about AI models"

/task list Show all scheduled tasks with status (PENDING, RUNNING, COMPLETED)

/task delete <id> Cancel a specific scheduled task by ID

/todo auto add "..." is the preferred unified format. /task exists for convenience.

`/channels`

Manage Telegram channel integration. Configure bot token, allowed users, and active channels.

Usage

list Show current Telegram config (bot token mask, user ID, chat ID)

set-bot-token /channels set-bot-token <123456:ABC-DEF...>

set-user-id /channels set-user-id <123456789> — who is allowed to use the bot

set-chat-id /channels set-chat-id <-1001234567890> — which chat to respond in

switch /channels switch main|telegram:12345 — switch the active conversation channel

disconnect Stop Telegram polling

Groups are supported with TELEGRAM_ALLOW_GROUPS in settings. Each Telegram chat gets its own conversation with inline approval buttons.

`/remind`

Create a reminder using natural language time parsing. The agent will send you a message when the time arrives.

Usage

/remind <message at time>

relative /remind "in 30 minutes call mom"

absolute /remind "tomorrow at 7am make breakfast"

day of week /remind "next monday at 9am standup notes"

Requires the agent to be running at the scheduled time. Reminders are stored in the cron manager and show up in /cron list.

`/cron`

Manage scheduled reminders — list active tasks, cancel specific ones, or clear completed reminders.

Usage

/cron <list|delete|clear>

list Show all pending, fired, or cancelled reminders with status icons and scheduled times

delete /cron delete <id> — cancel a specific reminder by ID

clear Remove all fired/completed reminders from the list

`/heartbeat`

Manage the autonomous heartbeat cycle — scheduled runs where the agent operates in yolo mode, reading memory files, executing tasks from HEARTBEAT.md, and reporting back.

Usage

status Show heartbeat status, active schedule, and all schedule configs

enable / disable Turn the heartbeat cycle on or off

intraday /heartbeat intraday 8:10,10:10,14:20 — specific times during the day (comma-separated H:MM)

daily /heartbeat daily 6:00 — run every day at a specific time

weekly /heartbeat weekly monday@6:00 — run every week on a specific day

monthly /heartbeat monthly 1@6:00 — run on day 1 of each month

dreaming /heartbeat dreaming 2:00 — set the late-night "dreaming" cycle time for reflective thinking

now Run the heartbeat immediately

The heartbeat reads MEMORY.md, USER.md, AGENTS.md, TODO.md and the active schedule section from HEARTBEAT.md, then executes tasks autonomously.

`/agent`

Launch an external AI agent subprocess. The agent runs in parallel while you continue working in the main session.

Usage

--mode Set approval mode for the spawned agent (plan, edit, auto, or yolo)

--effort Set reasoning effort for the spawned agent

prompt The task to delegate to the agent

Example

/agent --mode auto --effort medium refactor auth module

Output streams to the Agents tab in the TUI. Multiple agents can run concurrently.

`/init`

Run the interactive setup wizard — step-by-step configuration of provider, API key, model, and identity files.

Usage

/init

wizard steps 1) Choose provider (Anthropic, OpenAI, Google, local, Ollama, OpenRouter) → 2) Enter API key → 3) Select model → 4) Set identity (SOUL.md, USER.md, AGENTS.md, MEMORY.md, TOOLS.md)

If SOUL.md and settings.json already exist, the wizard assumes initialization is complete. Can also pass an API key directly: /init <sk-...>.

`/snapshots`

List recent git-stash snapshots for recovery. Snapshots are created automatically before each turn in yolo mode.

Usage

/snapshots

Returns a numbered list of snapshots with SHA, timestamp, and message. Use with /revert to restore a previous state. Silently disabled outside a git repository.

`/revert`

Revert the working directory to a git snapshot. Restores files to their state before a specific turn.

Usage

/revert [index]

index Snapshot index (0 = most recent, default: 0)

Works only with snapshots created by /snapshots. Restoring creates a new git commit that reverts files to the snapshot state.

`/exit`

Exit curie-agent and close the current session. The session is saved and can be resumed later with curie-agent resume or curie-agent continue.

Usage

/exit

Session data is preserved in ~/.curie-agent/sessions/<id>/ for later inspection or resumption.

`/wiki`

Interact with the compounding knowledge wiki from within the TUI. Search pages, view or edit a specific page, check wiki health, or explore the backlink graph.

Usage

/wiki [page <name>|search <query>|lint|graph]

/wiki Show wiki status: path, page count, index size, last activity

/wiki search /wiki search "prompt caching" — full-text search across all wiki pages

/wiki page /wiki page concepts/prompt-caching — display a specific wiki page in the TUI

/wiki lint Run health check: orphan pages, broken wikilinks, missing-from-index entries

/wiki graph Emit the backlink graph as JSON — nodes are pages, edges are [[wikilinks]]

The wiki lives at ~/.curie-agent/wiki/ by default. Override with wiki.path in settings. Scaffold with curie-agent wiki init first.

`/system`

Show system information: OS, platform, Node.js version, PathGuard status, and current working directory. Useful for debugging environment issues.

Usage

/system

In the web dashboard, this opens a system-info panel with the same details plus daemon version and WebSocket connection status.

`/model pricing`

View or configure tiered pricing for the active model. Supports multi-tier format with input, output, and cache prices per million tokens.

Usage

/model pricing [in;out;cache_write;cache_read]

no args Show current pricing configuration and estimated cost for the active session

set pricing /model pricing 3.0;15.0;3.75;0.3 — set input;output;cache_write;cache_read per million tokens

Pricing is used by the TokenMonitor to estimate session cost and trigger tier-crossing alerts. Default prices are loaded from the built-in cost reference table for known models.

`/run`

Execute a skill by name. Loads the skill's instruction file and activates it for the current session.

Usage

/run <skill-name>

skill-name Name of a skill in ~/.curie-agent/skills/ or .curie-agent/skills/

Use /skill to list available skill names before running. Skills can also be invoked by prompt — /run forces activation by name.

`/clear`

Clear the current conversation history from the chat surface. Does not end the session or affect memory files.

Usage

/clear

The session remains active — model, mode, and all settings are preserved. Use /context compact instead if you want to summarize before clearing.

`/copy`

Copy the last assistant response to the system clipboard.

Usage

/copy

Handy for pulling code blocks or formatted text out of the TUI without manually selecting text. Works in both TUI and web dashboard.

`/spawn`

Alias for /agent. Spawn a subagent that runs concurrently in a separate TurnLoop instance.

Usage

See /agent for full parameter documentation. Output streams to the Agents tab.