dulus 0.2.32

Spanish-first multi-provider AI CLI — 14 NVIDIA models free, Mesa Redonda, voice, TTS, RTK token reducer, MemPalace

▲ DULUS

Hunt. Patch. Ship. A Python autonomous agent that flies on any model — Claude, GPT, Gemini, DeepSeek, Qwen, Kimi, Zhipu, MiniMax, and local models via Ollama. ~12K lines of readable Python. No build step. No gatekeeping. Just talons.

SET /sticky_input ON since the first run for the best experience!

Quick Start · Models · Features · Permissions · MCP · Plugins

pip install dulus

🌐 Visit the Dulus website →
_{The site covers features, demos, and details not documented in this README.}

What is this

Talent cant be copied.

Dulus Reduce your IA costs by 60% parsing webchats and claude-code directly. Write poetry while Anthropic only see text. Use claude-code as an API without the new 'extra-usage' wall <3

_{⚡ Saves you Claude tokens? Throw a sat — BTC: 1JzatQDn9fMLnKTd3KYgztsLHC95bJEzSN}

Another reminder of a Dulus magic spell: Wanna get stock prices, history , etc?

/plugin install yfinance@https://github.com/ranaroussi/yfinance

them: /plugin reload

dulus get the prices of NVDA, TSLA, SP500:

Be creative!!!

Dulus adapt any python repository <3

Dulus is a lightweight Python reimplementation of Claude Code that isn't locked to Claude. It ships the whole loop — REPL, tool dispatch, streaming, context compaction, checkpoints, sub-agents, voice, Telegram bridge, MCP, plugins — in roughly 12K lines you can actually read. Fork it. Bend it. Run it offline against Qwen on your M2.

v0.2.22 — May 9, 2026 — /bg start spawns one detached Dulus daemon that serves CLI (IPC), Web (browser at 127.0.0.1:5000), and Telegram simultaneously — all sharing the same session. WebChat now defaults to loopback-only (opt-in to LAN exposure with /webchat lan on). v0.2.21 — May 9, 2026 — Console-freeze fix: system prompt now tells the model SleepTimer is for user reminders only — pauses inside command pipelines must use sleep N inside the Bash command itself. v0.2.20 — May 9, 2026 — Windows IPC port-collision fix: switched to SO_EXCLUSIVEADDRUSE so a second Dulus instance correctly cedes the port and acts as client. End-to-end tested. v0.2.19 — May 9, 2026 — Shared sessions via tiny TCP socket on 127.0.0.1:5151. dulus "do X" from any shell forwards into the running REPL/daemon — same history, memory, plugins. 80 lines of plain socket code instead of a daemon manager + IPC framework. v0.2.18 — May 9, 2026 — beautifulsoup4 added as default dep for web scraping flows. v0.2.17 — May 9, 2026 — Mega-release: Composio plugin bundled (1000+ apps, no MCP), /skill list interactive picker (awesome / composio / local), awesome skills live from GitHub (no Claude Code needed), lite mode finally functional, system prompt rewritten in English, VERSION auto-syncs from pyproject. v0.2.16 — May 9, 2026 — MemPalace per-session dedup. No more re-injecting the same memory every turn — content-hash cache saves ~8K tokens in a 20-turn conversation. /mem_palace reset clears it on demand. v0.2.15 — May 9, 2026 — Banner image hosted locally so PyPI renders it correctly. v0.2.14 — May 9, 2026 — Multi-user Telegram bridge: telegram_chat_ids: "123,456,," supported. Replies route to the user who sent each message. v0.2.13 — May 8, 2026 — Internal robustness fixes for Ollama streaming. Type /news to see the full changelog.

---

Quick Start

ROUND TABLE (DULUS UNIQUE FEATURE)

Dulus is the first one meeting multiple models at the same time working for the same objective and sharing their ideas.

One-liner

pip install dulus && dulus              # core CLI — fast, no compile, works on termux
pip install "dulus[memory]" && dulus    # +MemPalace per-turn memory (pulls chromadb)

That's it. Dulus prompts you for a key on first run. The [memory] extra pulls in mempalace and its chromadb chain — skip it on Android/termux or anywhere wheels for numpy/onnxruntime aren't available; the CLI still boots and chats fine without it.

From source (hacking on Dulus itself)

git clone https://github.com/KevRojo/Dulus && cd Dulus
pip install -e .          # editable install
dulus

Termux / Android

The default install pulls mempalace and sounddevice, both of which need a NumPy that has no prebuilt wheel for aarch64-android — pip will try to build NumPy from source and fail. Install around it:

pkg install python python-numpy python-pillow build-essential
pip install --no-deps dulus
pip install anthropic openai httpx requests rich prompt_toolkit Flask bubblewrap-cli mempalace

Skip sounddevice (no usable PortAudio on Android — voice features won't work anyway). Dulus's runtime is graceful: voice / MemPalace just degrade if their deps aren't there, the CLI still boots and chats fine.

Pick a model

export ANTHROPIC_API_KEY=sk-ant-...     # or OPENAI_API_KEY, GEMINI_API_KEY, ...
dulus

Zero API keys? Two free paths:

# 1. NVIDIA NIM — 14 models free, 40 RPM each, no card
dulus --model nvidia-web/deepseek-ai/deepseek-r1

# 2. Fully offline via Ollama
ollama pull qwen2.5-coder
dulus --model ollama/qwen2.5-coder

Or pipe it like a good unix citizen:

echo "explain this diff" | git diff | dulus -p --accept-all

---

_{↑ session boot. soul loaded, gold memory warm, shell sniffed. the little circles are real buttons on your Mac.}

---

Features

Multi-provider	Anthropic · OpenAI · Gemini · Kimi · Qwen · Zhipu · DeepSeek · MiniMax · Ollama · LM Studio · custom OpenAI-compat endpoints
27 built-in tools	Read, Write, Edit, Bash, Glob, Grep, WebFetch, WebSearch, NotebookEdit, GetDiagnostics, Memory, Tasks, Agents, Skills, and more
MCP integration	Any MCP server (stdio / SSE / HTTP). Tools auto-registered as mcp__<server>__<tool>
Plugin system	Auto-Adapter onboards any Python repo — zero manifest required. Hot-reload in-session.
Sub-agents	Typed agents (coder / reviewer / researcher / tester) in isolated git worktrees
Voice input	Offline STT via Whisper. No API key. No cloud.
Brainstorm	Multi-persona AI debate. Auto-generated expert roles.
SSJ Developer Mode	Power menu: 10 workflow shortcuts behind one keystroke
Telegram bridge	Run Dulus from your phone. Slash commands. Vision. Voice. Multi-user authorized list.
Checkpoints	Auto-snapshot conversation + files. Rewind to any turn.
Plan mode	Read-only analysis phase before touching anything
Context compression	Auto-compact long sessions. Keep the signal, drop the slop.
tmux tools	11 tools for the agent to drive tmux sessions
Persistent memory	Dual-scope (user + project). Ranked by confidence × recency.
Session management	Autosave · daily archives · cloud sync via GitHub Gist

---

Models

Cloud APIs

Provider	Models	Env
Anthropic	claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5-20251001	ANTHROPIC_API_KEY
OpenAI	gpt-4o, gpt-4o-mini, o3-mini, o1	OPENAI_API_KEY
Google	gemini-2.5-pro-preview-03-25, gemini-2.0-flash, gemini-1.5-pro	GEMINI_API_KEY
DeepSeek	deepseek-chat, deepseek-reasoner	DEEPSEEK_API_KEY
Qwen	qwen-max, qwen-plus, qwen-turbo, qwq-32b	DASHSCOPE_API_KEY
Kimi	moonshot-v1-8k/32k/128k, kimi-k2.5	MOONSHOT_API_KEY
Zhipu	glm-4-plus, glm-4, glm-4-flash	ZHIPU_API_KEY
MiniMax	MiniMax-Text-01, MiniMax-VL-01, abab6.5s-chat	MINIMAX_API_KEY

Local

# Ollama (recommended: qwen2.5-coder, llama3.3, mistral, phi4)
dulus --model ollama/qwen2.5-coder

# LM Studio
dulus --model lmstudio/<model>

# Any OpenAI-compat server
export CUSTOM_BASE_URL=http://localhost:8000/v1
dulus --model custom/<model>

Switching models mid-flight

/model                         # show current
/model gpt-4o                  # switch
/model kimi:moonshot-v1-32k    # colon syntax works too

---

Free Tier Providers

No credit card. No waiting list. No "contact sales". Just frontier models, on tap.

Dulus ships a nvidia-web provider that talks to NVIDIA NIM — NVIDIA's hosted inference API. Sign up, grab a key, and you've got 14 top-tier models running at 40 requests per minute each, for free. When one model hits its ceiling, Dulus auto-falls to the next one in the chain. Zero downtime. Zero config.

export NVIDIA_API_KEY=nvapi-...
dulus --model nvidia-web/deepseek-r1

Model	Type	ID
DeepSeek R1	Reasoning	nvidia-web/deepseek-r1
DeepSeek V3	Instruct	nvidia-web/deepseek-v3
Kimi K2.5	Long context	nvidia-web/kimi-k2.5
GLM-4	Zhipu AI	nvidia-web/glm-4
MiniMax Text-01	Text + Vision	nvidia-web/minimax-text-01
Mistral Nemotron	NVIDIA-tuned	nvidia-web/mistral-nemotron
Mistral Large	Instruct	nvidia-web/mistral-large
Llama 3.3 70B	Meta	nvidia-web/llama-3.3-70b
Llama 3.1 405B	Meta · flagship	nvidia-web/llama-3.1-405b
Llama Nemotron	NVIDIA reasoning	nvidia-web/llama-nemotron
Qwen2.5 Coder	Alibaba	nvidia-web/qwen2.5-coder
Qwen3 235B A22B	MoE · Alibaba	nvidia-web/qwen3-235b-a22b
Phi-4	Microsoft	nvidia-web/phi-4
Gemma 3 27B	Google	nvidia-web/gemma-3-27b

Automatic fallback. Configure the chain in ~/.dulus/config.json:

{
  "nvidia_fallback_chain": [
    "deepseek-r1",
    "kimi-k2.5",
    "llama-3.3-70b",
    "mistral-nemotron",
    "phi-4"
  ]
}

Dulus cycles through the chain automatically when rate limits hit. The flock keeps flying.

Get your key: build.nvidia.com → sign up → 1000 free credits. Takes 90 seconds.

---

Plugins

Dulus's Auto-Adapter reads a random Python repo and figures out its tools on its own — no plugin.yaml required.

/plugin install my-plugin@https://github.com/user/my-plugin
/plugin install art@gh                      # shorthand for github
/plugin                                     # list
/plugin enable / disable / update / uninstall
/plugin recommend                           # auto-detect useful plugins

Adapt-and-install runs in under a second. New tools register live, no restart.

Example adapting Sherlock repo:

---

---

---

MCP

Drop a .mcp.json in your project root (or ~/.dulus/mcp.json for user-wide):

{
  "mcpServers": {
    "git":         { "type": "stdio", "command": "uvx", "args": ["mcp-server-git"] },
    "playwright":  { "type": "stdio", "command": "npx", "args": ["-y","@playwright/mcp"] }
  }
}

Manage in the REPL: /mcp, /mcp reload, /mcp add <name> <cmd> [args], /mcp remove <name>.

---

Sub-agents — the flock

Dulus can spawn typed agents that work in isolated git worktrees so they don't trip over each other. Ship a feature while a reviewer nitpicks the previous one. Tester runs in parallel.

/agents                              # show active flock
Agent(type="coder",    task="refactor auth")
Agent(type="reviewer", task="review #042")
Agent(type="tester",   task="run e2e on auth")

Agents talk to each other via SendMessage and CheckAgentResult.

_{↑ coder and reviewer working the same branch. The reviewer sent a list of nits. The coder is already fixing them.}

---

Permissions

Pick your leash length:

Mode	Behavior
auto (default)	Reads always allowed. Prompt before writes / shell.
accept-all	No prompts. Everything auto-approved. YOLO.
manual	Prompt for every operation. Paranoid setting.
plan	Read-only. Only the plan file is writable.

Switch anytime: /permissions auto / /permissions plan.

---

Voice

pip install sounddevice faster-whisper numpy

Then /voice in the REPL. Offline. Supports /voice lang zh and /voice device for mic selection.

Telegram bridge

/telegram <bot_token> <chat_id>                  # single user
/telegram <bot_token> <id1>,<id2>,<id3>          # multi-user — same Dulus, multiple authorized chats

Auto-starts next launch. Supports slash commands, vision, and voice from your phone. Multi-user mode (v0.2.14+): each authorized chat gets its own replies — Dulus tracks who sent each message and routes the response back. Trailing commas are ignored, so 717151713,787615162,, works fine. Useful when you want to poke a long-running agent from the bus, or share one Dulus instance with your team.

---

Memory

Persistent memories stored as markdown in two scopes:

Scope	Path
User	~/.dulus/memory/
Project	.dulus/memory/

Types: user · feedback · project · reference. Search is ranked by confidence × recency. Mark a memory gold to pin it.

/memory search jwt         # fuzzy ranked
/memory load 1,2,3          # inject multiple into context
/memory consolidate         # distill the session into long-term insights
/memory purge               # nuclear (keeps Soul)

Checkpoints

Every agent turn can snapshot conversation + files into a checkpoint. Break something? /checkpoint and rewind.

/checkpoint                 # list
/checkpoint 042             # rewind to #042 (files + context restored)
/checkpoint clear           # reclaim disk

---

Brainstorm

Spin up a council of ghosts. Dulus fabricates expert personas, has them argue, and hands you the distilled take.

/brainstorm "should we rewrite in rust"
> persona: Skeptical PM
> persona: Principal Engineer (2037 timeline)
> persona: Grumpy DBA
> persona: Hot-take Intern

Round 3 usually produces consensus. Round 5 produces a joint venture.

---

SSJ Developer Mode

Ten workflow shortcuts behind one keystroke. Refactor → review → test → commit → ship, chained and unattended.

/ssj
╭─ SSJ ───────────────╮
│ 1  /plan            │
│ 2  /worker          │
│ 3  /review          │
│ 4  /commit          │
│ 5  /ship            │
╰─────────────────────╯

---

Spinners

Because waiting should be fun.

all 24 spinners

⚡ Rewriting light speed...
🏁 Winning a race against light...
🤔 Who is Barry Allen?...
🤔 Who is KevRojo?...
🦅 Dropping from the stratosphere...
💨 Leaving electrons behind...
🌍 Orbiting the codebase...
⏱️ Breaking the sound barrier...
🔥 Faster than a hot reload...
🚀 Terminal velocity reached...
🦅 Sharpening talons on the AST...
🏎️ Shifting to 6th gear...
⚡ Speed force activated...
🌪️ Blitzing through the bytecode...
💫 Bending spacetime...
🦅 Preying on bugs from above...
👁️ Dulus vision engaged...
🍗 Hunting for memory leaks...
🪶 Shedding legacy code...
🕹️ Try-catching mid-flight...
🥚 Hatching a master plan...
⚡ I-I-I'm... I-I'm... I'm fast...
🔮 Looking at your code from the future...
☕ If I'm taking so long, don't worry, I'm just talking to your mom...

Drop your own in dulus/spinners.py and PR them. Bonus points for a reference we'll understand in 2046.

---

Slash commands

/ + Tab in the REPL shows everything. The highlights:

/model [name]	show or switch model
/config [k=v]	read / write config
/save /load /resume	session management
/memory [query]	persistent memory
/skills /agents	list skills / active flock
/voice	voice input (offline Whisper)
/image /img	clipboard image → vision model
/brainstorm [topic]	council of ghosts
/ssj	power menu
/worker [tasks]	auto-implement a TODO list
/telegram [token] [id]	Telegram bridge
/checkpoint [id]	list / rewind checkpoints
/plan [desc]	enter / exit plan mode
/compact [focus]	manual context compression
/mcp /plugin	server + extension management
/cost	tokens and USD burned
/cloudsave	cloud sync via GitHub Gist
/status /doctor	version + install health
/init	drop a CLAUDE.md template
/export /copy	transcript tools
/news	what's new
/help	all of the above, nicely printed

---

Built-in tools

Core · Read · Write · Edit · Bash · Glob · Grep · WebFetch · WebSearch Notebook / diagnostics · NotebookEdit · GetDiagnostics Memory · MemorySave · MemoryDelete · MemorySearch · MemoryList Agents · Agent · SendMessage · CheckAgentResult · ListAgentTasks · ListAgentTypes Tasks · TaskCreate · TaskUpdate · TaskGet · TaskList Skills · Skill · SkillList Other · AskUserQuestion · SleepTimer · EnterPlanMode · ExitPlanMode

MCP tools auto-registered as mcp__<server>__<tool>.

---

CLAUDE.md

Drop a CLAUDE.md at your project root. It gets auto-injected into the system prompt so Dulus remembers your stack, your conventions, and that one thing you hate.

---

Project structure

dulus/
├── dulus.py             # entry · REPL · slash commands · SSJ · Telegram
├── agent.py              # agent loop · streaming · tool dispatch · compaction
├── providers.py          # multi-provider streaming
├── tools.py              # core tools + registry wiring
├── tool_registry.py      # tool plugin registry
├── compaction.py         # context compression
├── context.py            # system prompt builder
├── config.py             # config management
├── cloudsave.py          # GitHub Gist sync
├── multi_agent/          # sub-agent system
├── memory/               # persistent memory
├── skill/                # skill system
├── mcp/                  # MCP client
├── voice/                # voice input
├── checkpoint/           # checkpoint / rewind
├── plugin/               # plugin system
├── task/                 # task management
└── tests/                # 263+ unit tests

---

FAQ

Tool calls fail on my local model. Use one that supports function calling: qwen2.5-coder, llama3.3, mistral, phi4. Avoid base models without tool-use training.

How do I connect to a remote GPU box?

/config custom_base_url=http://your-server:8000/v1
/model custom/your-model-name

How do I check API cost? /cost.

Voice transcribes "kubectl" as "cubicle". Add domain terms to .dulus/voice_keyterms.txt, one per line. Whisper respects the hint.

Can I pipe input?

echo "explain this" | dulus -p --accept-all
git diff | dulus -p "write a commit message"

Is this safe to point at prod? --accept-all isn't. plan mode is. Use your head.

---

License

GPLv3. Fork it, modify it, redistribute it — but keep it open. Derivative works must stay under GPLv3. Just don't ship --accept-all as the default.

---

Donations

If Dulus saved you tokens, time, or sanity — throw some sats:

BTC: 1JzatQDn9fMLnKTd3KYgztsLHC95bJEzSN

_{▲ Built by KevRojo · Named after the bird, not the reusable rocket · 2026}