5 Open‑Source AI Projects: Lark CLI, OpenSpace, G0DM0D3, Awesome‑AI List, and Meta TribeV2

1. larksuite/cli – Lark Official CLI Tool

Stars: 3,716. The CLI covers almost all Feishu business modules (messages, calendar, docs, sheets, knowledge base, mail, tasks, meetings) with more than 200 commands and 19 built‑in AI Agent Skills. Each command returns structured JSON, enabling both users and AI agents to operate Feishu via the command line.

# Send a message
lark-cli im +messages-send --chat-id "oc_xxx" --text "Meeting started"

# View calendar
lark-cli calendar +agenda

# Create a document
lark-cli doc +create --title "Meeting minutes"

Key highlights:

Coverage of 11 business domains with 200+ commands.

AI‑friendly output (structured JSON) and concise parameters.

Three‑minute onboarding: npm install → configure app credentials → login.

Identity switching between user and bot.

Credentials stored in system Keychain; input‑injection protection; terminal output sanitization.

Repository:

https://github.com/larksuite/cli

2. HKUDS/OpenSpace – Self‑Evolving AI Agents

Stars: 2,532. OpenSpace addresses the limitation that current AI agents do not learn from real‑world experience and cannot share knowledge.

Core capabilities (plug‑in style for agents such as OpenClaw, nanobot, Claude Code, Codex, Cursor, etc.):

Automatic repair: broken skills are fixed automatically.

Automatic improvement: successful task patterns are upgraded to better skill versions.

Automatic learning: effective workflows are captured from actual usage.

Cloud‑based Skill Community enables any agent to share learned skills with optional public, private, or team access controls.

Token‑saving benchmark (GDPVal, 50 professional tasks): OpenSpace agents achieved a 4.2× economic gain while reducing token consumption by 46% compared with baseline agents.

Highlights:

Collective intelligence: one agent’s learning benefits all agents.

4.2× economic benefit + 46% token reduction.

Plugin‑style integration with major AI agents.

Continuous evolution via the cloud community, eliminating the need for retraining from scratch.

Repository:

https://github.com/HKUDS/OpenSpace

3. elder‑plinius/G0DM0D3 – Open‑Source Multi‑Model Chat Interface

Stars: 2,040. G0DM0D3 ("LIBERATED AI") runs as a single index.html file in the browser, requiring no login or registration.

Integrates OpenRouter to support 50+ models (Claude, GPT‑5, Gemini, Grok, Mistral, LLaMA, DeepSeek, Qwen, etc.). Built‑in modes:

GODMODE CLASSIC – Parallel racing of five vetted model‑prompt combos; the best answer wins.

ULTRAPLINIAN – Multi‑model evaluation engine with five tiers (Fast/Standard/Smart/Power/Ultra), evaluating up to 55 models concurrently.

Parseltongue – Input‑perturbation engine for robustness testing, offering 33 attack techniques across three intensity levels.

AutoTune – Automatic hyper‑parameter tuning that adapts sampling settings (temperature, top_p, etc.) based on question type and learns from thumbs‑up/down feedback.

Privacy guarantees:

No login required.

API keys stored only in browser localStorage, never uploaded.

No cookies or tracking.

Telemetry can be disabled; all collected data is open‑source and auditable.

Repository:

https://github.com/elder-plinius/G0DM0D3

4. alvinunreal/awesome‑opensource‑ai – Curated List of Truly Open AI Projects

Stars: 1,949. The repository curates projects that fully open all core components: model, training code, training data, and logs.

Entries are classified into 14 categories, including:

Core frameworks (PyTorch, Hugging Face Transformers).

Open models (Qwen 3.5, DeepSeek‑V3.2/R1, Gemma 3, MiniMax‑M2, Kimi K2.5, OLMo 2).

Inference engines (vLLM, TensorRT‑LLM, Ollama).

Agents and multi‑agent systems.

RAG and knowledge systems.

Generative media tools.

Fine‑tuning toolkits (LLaMA‑Factory, Axolotl, Unsloth).

MLOps/LLMOps.

Evaluation & benchmarks.

AI safety & alignment.

User interfaces & self‑hosted platforms.

Developer tools.

Highlights:

Strict selection ensures every listed project is genuinely open.

Clear categorization covers the entire AI stack from models to applications.

Each entry displays a GitHub Stars badge for quick navigation.

All projects are released under CC0, allowing unrestricted use.

Repository:

https://github.com/alvinunreal/awesome-opensource-ai

5. facebookresearch/tribev2 – Multimodal Brain‑Encoding Model

Stars: 931. TribeV2 predicts fMRI responses to natural video, audio, and text stimuli. It combines three state‑of‑the‑art feature extractors:

LLaMA 3.2 for text understanding.

V‑JEPA2 for video understanding.

Wav2Vec‑BERT for audio understanding.

These encoders feed into a unified Transformer that maps representations onto the cortical surface.

from tribev2 import TribeModel

model = TribeModel.from_pretrained("facebook/tribev2", cache_folder="./cache")

df = model.get_events_dataframe(video_path="path/to/video.mp4")
preds, segments = model.predict(events=df)
print(preds.shape)  # (time steps, cortical vertices)

Inputs can be video, audio, or text (automatic speech‑to‑text conversion). Outputs are predicted fMRI brain responses, visualizable on cortical maps.

Research value:

Helps neuroscientists understand multimodal information processing in the brain.

Advances cross‑disciplinary work in computational neuroscience and AI interpretability.

Validated on public datasets such as Algonauts2025.

Highlights:

First foundation model covering vision, audition, and language modalities.

Trained on real fMRI data, offering strong interpretability.

Provides a ready‑to‑run Colab notebook for one‑click experimentation.

Model weights are openly available on HuggingFace.

Repository:

https://github.com/facebookresearch/tribev2