Four Pillars
One binary, one config, one TUI. Everything you need, nothing you don't.
Everyday Assistant
Telegram gateway · heartbeat routines · cron reminders · 6 providers. Morning briefings on your phone, inline approval buttons, per-chat sessions and group support.
Coding Agent
Turn loop · 14 tools · MCP client · 4 approval modes · safety guards. Read, Edit, Write, Glob, Grep, Bash, Wiki — with path guard, command guard, and git snapshots.
Compounding Wiki
Markdown-on-disk knowledge base that compounds as you work. Ingest articles, PDFs, and URLs — the agent writes summary and entity pages, updates the index, and answers "how do I X" with cited wiki pages.
Multi-Session Orchestrator
JSON-RPC daemon (40+ RPC methods, WebSocket, bearer-token auth) ships. tmux-like pane grid, broadcast, queue, A/B diff, and YAML playbooks coming in Phase 5.
Your agent knows you
Five identity files live in ~/.curie-agent/. Created by /init, edited by you, read at every session start.
SOUL.md
Customizable
Your agent's personality, voice, and operating principles. Define the archetype, tone, and behavioral constraints that make it yours.
USER.md
Customizable
Your profile: name, timezone, goals, expertise, preferences. The agent reads this to tailor every response — so you never repeat yourself.
MEMORY.md
Auto-updated
Accumulating knowledge — facts, heuristics, project history, lessons learned. Grows with every heartbeat synthesis cycle.
AGENTS.md
Customizable
Agent registry and role definitions. Injected into the context at every session start to orient the agent for your workspace.
HEARTBEAT.md
Auto-updated
Proactive routine definitions: morning briefing, daily synthesis, nightly dreaming. Configured via /heartbeat commands.
Ready Now
Every feature below ships with v0.3.0. No config required — just install, run /init, and go.
6 AI Providers
Anthropic, OpenAI, Google Gemini, Ollama, Local (OpenAI-compatible), OpenRouter. Switch at runtime with /provider.
14 Built-in Tools
Read, Edit, Write, Glob, Grep, Bash, WebSearch, WebFetch, Reminder, Skill, Spawn Agent, Task, Todo, Wiki — all with Zod-validated schemas.
4 Approval Modes
Plan (read-only, writes only under plans/), Edit (auto-edits, asks shell), Auto (LLM harm-check per call), Yolo (no prompts).
32+ Slash Commands
/model, /theme, /effort, /tools, /websearch, /remind, /cron, /context, /stats, /mcp, /heartbeat, /todo, /task, /skill, /wiki, /run, /clear, /copy, /spawn, /system, /snapshots, /revert, /exit, and more.
MCP Client
Connect to MCP servers via stdio, SSE, or streamable-HTTP. Auto-discovers tools. Manage with /mcp commands.
Telegram Gateway
Control your agent from Telegram. Inline approval buttons, per-chat sessions, group support.
Interactive Init Wizard
Step-by-step setup: choose provider, set API key, pick model, create identity files (SOUL.md, USER.md, AGENTS.md).
Cron & Reminders
Natural language reminders ("remind me in 30 min"), scheduled cron tasks, heartbeat system for autonomous cycles.
8 Themes + 6 TUI Tabs
Tokyo Night, Nord, Dracula, Solarized, Gruvbox, Black, White, Grey. Chat, Channels, Stats, Projects, Agents, Wiki.
Git Snapshots
Before each turn in yolo mode, a dangling git commit is created. /snapshots lists them, /revert restores. Never lose work again.
Daemon Protocol
JSON-RPC over WebSocket for remote control. Auth, channel management, and slash commands via programmatic API — headless anywhere.
Identity System
SOUL.md + USER.md + MEMORY.md. Your agent learns who you are and remembers it across every session. Fully customizable markdown files.
Built-in Skills
Deep Research (4-phase methodology, citations) and Planning (7 modes: roadmap, timeline, goals, habits...). Drop your own SKILL.md to extend.
Context Management
Auto-compaction at configurable threshold (default 75%). /context visual fill grid. Manual compact with /context compact.
Compounding Wiki
Markdown-on-disk knowledge base. Ingest articles, PDFs, URLs — agent writes summary and entity pages, updates the index, answers questions with cited pages.
Subagents
Spawn in-process subagents via /agent or spawn_agent tool. Each runs its own TurnLoop concurrently — visible in the TUI Agents tab and web SubagentsView.
Thinking Streaming
Watch the model reason in real time. Extended thinking tokens stream as a collapsible block — toggle visibility with Ctrl+O in the TUI.
Your Agent, On a Schedule
Set up autonomous heartbeat cycles — the agent checks in, reviews your workspace, and updates your TODO.md while you sleep.
Heartbeat System
Shipped- Reads MEMORY.md, USER.md, AGENTS.md, TODO.md
- Updates project status and tracks progress
- Runs in yolo mode — no approval prompts
- Delivers results to configured channels
A day with curie-agent
Built-in skills
Activate with a prompt or slash command. Drop your own SKILL.md in ~/.curie-agent/skills/<name>/ to extend.
/deep-research
Multi-source structured research with citations, source evaluation, and synthesis. Four-phase methodology: Orientation → Targeted Deep → Verification → Gap-fill.
/planning
Universal planning tool with 7 modes: Roadmap, Timeline, Goal Decomposition, Project, Periodic, Event, and Habit planning. Gathers context before committing.
wiki init / ingest / query
Persistent, compounding knowledge base. Ingest sources → agent writes pages. Query with natural language → cited answers. Lint for orphans and broken links. Stored in ~/.curie-agent/wiki/.
SKILL.md
Drop a SKILL.md in ~/.curie-agent/skills/<name>/. Compatible with Claude Code skills format. Use /skill to list and inspect available skills.
Roadmap
Phases 0–4 are complete. Wiki engine, Skills Runtime, Subagents, and MCP Client shipped.
Hooks & Pane Grid
Daemon (JSON-RPC + WebSocket) shipped. Pre/post hooks for tool calls (ToolUse, Stop, Compact, SessionStart), tmux-like pane grid, and YAML playbooks coming.
MCP Server
Expose curie-agent tools, sessions, wiki, and channels to other AI agents via MCP.
Subagents
In-process SubagentExecutor spawning concurrent TurnLoop instances. TUI Agents tab, Web SubagentsView, spawn_agent tool, daemon RPC.
MCP Client
Connect to MCP servers via stdio, SSE, or streamable-HTTP. Auto-discovers tools.
Skills Runtime
Claude-Code-compatible SKILL.md files in ~/.curie-agent/skills/ and ./.curie-agent/skills/.
Compounding Wiki
Markdown-on-disk knowledge base with Ingest, Query, and Lint workflows. Answers "how do I X" with cited wiki pages.
Your Provider, Your Choice
No vendor lock-in. Switch models and providers at runtime with a single slash command.
Add your own: @curie-agent/providers is a pluggable interface.
Up and Running in 30 Seconds
1. Install
npm install -g @curie-agent/cli2. Launch
curie-agent3. Configure
> /init4. Ask anything
> write me a TypeScript utility for retry logicretry.ts file with the implementation.Requires Node.js ≥ 20. Set ANTHROPIC_API_KEY or OPENAI_API_KEY to begin.