This server exposes the Arabic AI Atlas — a curated, offline catalogue of Arabic AI models, datasets, benchmarks, tools, agent skills and organizations with Hugging Face download metrics — through three read-only tools.
search — case-insensitive substring search across name, org, notes, tasks and tags, with exact filters for type (llm, asr, tts, ocr, embedding, dataset, benchmark, tool, agent-skill, org), country (SA, AE, EG, QA, MA, JO, TN, LB, KW, OM, BH, INTL) and modality (text, speech, vision, multimodal, none); results sorted by downloads, limit defaults to 10.
recommend — ranked, explained recommendations for a task (e.g. chat, tts, asr, ocr, embedding, translation); each result includes score and why, with optional dialect (msa, egy, gulf, lev, magh, iraqi, sudanese, yemeni, classical, mixed), on_device (phone/laptop CPU fit), license_filter ("open" or an exact license) and type; limit defaults to 3.
get — fetch one full entry by id (e.g. "jais-30b") including links and metrics, returning an error object for unknown ids.
Together these support discovery, filtering, comparison and selection of Arabic AI resources, and finding ids to inspect in detail.
Provides access to a curated catalogue of Arabic AI models, datasets and benchmarks sourced from Hugging Face, including download metrics; search and recommendation tools return entries with their Hugging Face model links and popularity stats.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@arabic-ai-atlas find an open-licensed Arabic speech recognition model for the Gulf dialect"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Open the interactive map → · search, filter by dialect or license, share a filtered link.
The Arabic AI ecosystem as a map, a list, and a skill your agent can install.
claude plugin marketplace add h9-tec/arabic-ai-atlas
claude plugin install arabic-ai-atlas@arabic-ai-atlasRequires uv on PATH.
Then ask Claude: which open Arabic TTS runs on a phone? It answers from this atlas via the recommend MCP tool.
Add this to your .mcp.json:
{"mcpServers": {"arabic-ai-atlas": {"command": "uv", "args": ["run", "--frozen", "--no-dev", "--directory", "/path/to/arabic-ai-atlas", "python", "mcp/server.py"]}}}Related MCP server: agent101-mcp
YAML in data/ is the only hand-edited source.
CI validates every PR against data/schema.json.
A nightly job pulls Hugging Face downloads and regenerates the map, README, dist/atlas.json and dist/llms.txt.
Wanted gaps (data/wanted.yaml) are re-evaluated on every build, and base_model links build the family tree and the MCP lineage tool.
The repo is itself a Claude Code plugin: 5 skills + an MCP server that reads the atlas offline.
Country | Entries | LLM | ASR | TTS | OCR | Embedding | Datasets | Tools | Benchmarks | Papers | Orgs |
2200 | 210 | 119 | 58 | 32 | 30 | 710 | 238 | 75 | 671 | 23 | |
🇸🇦 Saudi Arabia | 355 | 46 | 11 | 8 | 9 | 28 | 163 | 19 | 25 | 11 | 35 |
184 | 49 | 5 | 6 | 0 | 0 | 43 | 12 | 31 | 17 | 21 | |
🇪🇬 Egypt | 175 | 33 | 22 | 6 | 5 | 7 | 73 | 5 | 3 | 2 | 18 |
🇶🇦 Qatar | 128 | 9 | 3 | 0 | 2 | 0 | 63 | 10 | 17 | 19 | 5 |
🇲🇦 Morocco | 71 | 13 | 5 | 2 | 2 | 1 | 29 | 4 | 7 | 3 | 5 |
🇩🇿 Algeria | 42 | 3 | 1 | 3 | 0 | 0 | 18 | 15 | 1 | 0 | 1 |
🇵🇸 Palestine | 30 | 3 | 0 | 0 | 0 | 0 | 17 | 6 | 1 | 1 | 2 |
🇯🇴 Jordan | 25 | 1 | 1 | 0 | 1 | 0 | 15 | 2 | 2 | 1 | 2 |
🇱🇧 Lebanon | 23 | 14 | 0 | 0 | 0 | 0 | 6 | 1 | 0 | 1 | 1 |
🇹🇳 Tunisia | 22 | 2 | 1 | 1 | 0 | 0 | 11 | 0 | 1 | 3 | 3 |
🇾🇪 Yemen | 9 | 1 | 2 | 0 | 0 | 0 | 4 | 0 | 1 | 1 | 0 |
🇮🇶 Iraq | 8 | 2 | 0 | 0 | 0 | 0 | 4 | 0 | 1 | 0 | 1 |
🇴🇲 Oman | 8 | 2 | 0 | 0 | 0 | 0 | 2 | 1 | 0 | 0 | 3 |
🇸🇩 Sudan | 7 | 0 | 1 | 0 | 0 | 0 | 3 | 0 | 0 | 1 | 2 |
🇧🇭 Bahrain | 5 | 1 | 2 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 1 |
🇰🇼 Kuwait | 4 | 1 | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 1 | 1 |
🇱🇾 Libya | 3 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 |
🇸🇾 Syria | 3 | 2 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
The same atlas as a country-by-type grid, with the most downloaded entries named in each cell.
Showing 20 of 393 · see all 393 on the interactive map · full table
Name | Org | Country | Size | License | ⬇ Downloads | Updated | Links |
AraBERTv02 | AUB MIND Lab | 🇱🇧 LB | — | unknown | 558K | 2024-03-26 | |
Phi-4 | Microsoft | 🌍 INTL | 4B | mit | 357K | 2026-07-14 | |
gemma4 e4b claims comparison | k-chirkunov | 🌍 INTL | — | gemma | 342K | 2026-08-05 | |
Llama 3.3 | Meta | 🌍 INTL | 70B | llama3.3 | 257K | 2024-12-21 | |
Fanar-2-27B-Instruct | QCRI | 🇶🇦 QA | 27B | apache-2.0 | 205K | 2026-03-25 |
|
AraPoemBERT | King Saud University | 🇸🇦 SA | — | cc-by-nc-4.0 | 108K | 2024-05-23 | |
opus-mt-ar-en | Helsinki-NLP | 🌍 INTL | — | apache-2.0 | 106K | 2023-08-16 | |
bert-base-arabic-camelbert-mix-sentiment | CAMeL Lab, NYUAD | 🇦🇪 AE | — | apache-2.0 | 98K | 2021-10-17 | |
SentimentArEng | qandos0 | 🌍 INTL | — | unknown | 95K | 2023-12-14 | |
bert-base-arabic-camelbert-da-sentiment | CAMeL Lab, NYUAD | 🇦🇪 AE | — | apache-2.0 | 83K | 2021-10-17 | |
MARBERTv2 | UBC-NLP | 🌍 INTL | — | unknown | 65K | 2022-03-30 | |
opus mt tc big ar en | Helsinki-NLP | 🌍 INTL | — | cc-by-4.0 | 37K | 2023-08-16 | |
opus-mt-en-ar | Helsinki-NLP | 🌍 INTL | — | apache-2.0 | 30K | 2023-08-16 | |
bert-base-arabic-camelbert-msa-ner | CAMeL Lab, NYUAD | 🇦🇪 AE | — | apache-2.0 | 17K | 2021-10-17 | |
aragpt2-base | AUB MIND Lab | 🇱🇧 LB | 148M | unknown | 13K | 2023-10-30 | |
Fanar-1-9B | QCRI | 🇶🇦 QA | 9B | apache-2.0 | 13K | 2025-07-15 | |
bert-base-arabertv2 | AUB MIND Lab | 🇱🇧 LB | 136M | unknown | 12K | 2023-08-03 | |
bert-base-arabic-camelbert-mix-ner | CAMeL Lab, NYUAD | 🇦🇪 AE | — | apache-2.0 | 11K | 2021-10-17 | |
Falcon Arabic | TII (UAE) | 🇦🇪 AE | 7B | falcon-llm-license | 11K | 2025-05-31 | |
ALLaM | SDAIA & IBM | 🇸🇦 SA | 7B | apache-2.0 | 11K | 2025-07-14 |
Showing 20 of 174 · see all 174 on the interactive map · full table
Name | Org | Country | License | ⬇ Downloads | Updated | Links |
whisper-large-v3-turbo | OpenAI | 🌍 INTL | mit | 6.3M | 2024-10-04 | |
openai/whisper-large-v3 | OpenAI | 🌍 INTL | apache-2.0 | 3.8M | 2024-08-12 | |
wav2vec2-large-xlsr-53-arabic | jonatasgrosman | 🌍 INTL | apache-2.0 | 1.9M | 2022-12-14 | |
SeamlessM4T v2 | Meta | 🌍 INTL | cc-by-nc-4.0 | 305K | 2024-01-04 | |
MMS-1b-all | Meta | 🌍 INTL | cc-by-nc-4.0 | 277K | 2023-06-15 | |
Voxtral Mini | Mistral AI | 🌍 INTL | apache-2.0 | 174K | 2025-07-28 | |
cohere-transcribe-arabic-07-2026 | Cohere Labs | 🌍 INTL | apache-2.0 | 100K | 2026-07-13 | |
whisper-large-v3-turbo-ar-quran | Naazim | 🌍 INTL | apache-2.0 | 60K | 2025-12-08 | |
qwen3-asr-arabic-uae | Vadim Belsky | 🇦🇪 AE | apache-2.0 | 42K | 2026-04-04 | |
Whisper Quran | Tarteel AI | 🌍 INTL | apache-2.0 | 26K | 2022-12-13 | |
muaalem model v3 2 | obadx | 🌍 INTL | mit | 25K | 2025-09-04 | |
QwenCleo-ASR | Mohammed Aly | 🇪🇬 EG | apache-2.0 | 16K | 2026-06-17 | |
wav2vec2 quran phonetics | TBOGamer22 | 🌍 INTL | apache-2.0 | 11K | 2026-01-21 | |
wav2vec2 large xlsr 53 arabic quran v final | rabah2026 | 🌍 INTL | apache-2.0 | 9K | 2025-12-21 | |
egyptian-arabic-wav2vec2-xlsr-53 | Ibrahim Amin | 🇪🇬 EG | apache-2.0 | 5K | 2025-05-11 | |
Audar ASR V1 Turbo | Audar AI | 🌍 INTL | other | 4K | 2026-08-20 | |
Cohere-Speech-Tashkeel-2B | NAMAA-Space | 🇸🇦 SA | apache-2.0 | 2K | 2026-07-13 | |
nemotron 3.5 quran dual v5 | tamm5y5m5 | 🌍 INTL | unknown | 2K | 2026-09-02 | |
wav2vec2-large-xlsr-moroccan-darija | Boumehdi | 🇲🇦 MA | apache-2.0 | 2K | 2024-04-12 | |
whisper-large-v3-turbo-darija | Anas Zil | 🇲🇦 MA | mit | 2K | 2025-11-09 |
Showing 20 of 85 · see all 85 on the interactive map · full table
Name | Org | Country | License | ⬇ Downloads | Updated | Links |
XTTS-v2 | Coqui | 🌍 INTL | coqui-public-model-license | 6.7M | 2023-12-11 | |
Multilingual Chatterbox | Resemble AI | 🌍 INTL | mit | 1.7M | 2026-06-10 | |
f5tts-algerian-darja | Touati Kamel | 🇩🇿 DZ | apache-2.0 | 12K | 2026-09-23 | |
facebook/mms-tts-ara | Meta | 🌍 INTL | cc-by-nc-4.0 | 6K | 2023-09-01 | |
arabic-emirati-female-piper | Vadim Belsky | 🇦🇪 AE | mit | 1K | 2025-12-01 | |
speecht5_tts_clartts_ar | MBZUAI | 🇦🇪 AE | cc-by-nc-4.0 | 1K | 2025-09-10 | |
Nabra-82M-v0.1 | oddadmix | 🇪🇬 EG | apache-2.0 | 1K | 2026-07-03 | |
Audar TTS V1 Flash | Audar AI | 🌍 INTL | other | 1K | 2026-07-07 | |
DarijaTTS-v0.1-500M | Kandir Research | 🇲🇦 MA | apache-2.0 | 936 | 2025-11-26 | |
NAMAA-Saudi-TTS-V2 | NAMAA-Space | 🇸🇦 SA | cc-by-nc-sa-4.0 | 922 | 2026-04-21 | |
Audar TTS V1 Turbo | Audar AI | 🌍 INTL | other | 850 | 2026-07-07 | |
Arabic-TTS-Spark | IbrahimSalah | 🌍 INTL | fair-noncommercial-research-license | 830 | 2025-11-27 | |
Lahgtna OmniVoice v2 | oddadmix | 🇪🇬 EG | unknown | 554 | 2026-06-15 | |
habibi-tts-doda-darija | Jip7e | 🇲🇦 MA | apache-2.0 | 510 | 2026-08-29 | |
Hadra-TTS-f5 | Algerian NLP | 🇩🇿 DZ | apache-2.0 | 441 | 2026-09-23 | |
SILMA TTS v1 | SILMA AI | 🌍 INTL | apache-2.0 | 437 | 2026-08-23 |
|
Nabra 7M Distill | oddadmix | 🇪🇬 EG | apache-2.0 | 297 | 2026-09-16 | |
qwen3.5 TTS Emirati | Vadim Belsky | 🇦🇪 AE | apache-2.0 | 282 | 2026-03-15 | |
qwen3 TTS KSA | Vadim Belsky | 🇦🇪 AE | apache-2.0 | 262 | 2026-03-17 | |
voho-saudi-speak-0.6b | Voho AI | 🇸🇦 SA | cc-by-nc-sa-4.0 | 262 | 2026-09-14 |
Showing 20 of 51 · see all 51 on the interactive map · full table
Name | Org | Country | License | ⬇ Downloads | Updated | Links |
Arabic-handwritten-OCR-4bit-Qwen2.5-VL-3B-v2 | sherif1313 | 🌍 INTL | apache-2.0 | 8K | 2025-12-29 | |
arabic_PP-OCRv5_mobile_rec | PaddlePaddle | 🌍 INTL | apache-2.0 | 5K | 2025-10-16 | |
Qari-OCR-v0.3-VL-2B-Instruct | NAMAA-Space | 🇸🇦 SA | apache-2.0 | 2K | 2025-06-10 | |
Arabic-English-handwritten-OCR-v3 | sherif1313 | 🌍 INTL | apache-2.0 | 1K | 2025-12-29 | |
Qari OCR 0.2.2.1 VL 2B Instruct | NAMAA-Space | 🇸🇦 SA | apache-2.0 | 1K | 2025-06-07 | |
Qari-OCR-0.4.0-VL-4B-Instruct | NAMAA-Space | 🇸🇦 SA | apache-2.0 | 1K | 2026-02-01 | |
arabic PP OCRv3 mobile rec | PaddlePaddle | 🌍 INTL | apache-2.0 | 500 | 2025-07-22 | |
Qwen3-VL-2B-Persian-Arabic-Ocr-v1.0 | Mohajesmaeili | 🌍 INTL | apache-2.0 | 500 | 2025-12-23 | |
arabic-legal-documents-ocr-1.0 | Bakrianoo | 🇪🇬 EG | gemma | 486 | 2026-02-04 | |
Baseer Nakba | Misraj AI | 🇸🇦 SA | cc-by-nc-sa-4.0 | 469 | 2026-08-18 |
|
waqf ocr hand written v1 | Waqf AI | 🇪🇬 EG | apache-2.0 | 380 | 2026-06-09 | |
Arabic-Qwen3.5-OCR-v4 | sherif1313 | 🌍 INTL | apache-2.0 | 336 | 2026-03-24 | |
Arabic-GLM-OCR-v2 | sherif1313 | 🌍 INTL | apache-2.0 | 331 | 2026-05-27 | |
Mubsir Qwen 2B VL | Hatim2221 | 🌍 INTL | apache-2.0 | 315 | 2026-08-24 | |
Ketaba-OCR-LoRA | HassanB4 | 🌍 INTL | apache-2.0 | 304 | 2026-03-05 | |
Katib-Qwen3.5-0.8B-0.1 | oddadmix | 🇪🇬 EG | apache-2.0 | 270 | 2026-04-17 | |
Arabic GLM OCR v1 | sherif1313 | 🌍 INTL | apache-2.0 | 230 | 2026-03-29 | |
Qari-OCR-0.1-VL-2B-Instruct | NAMAA-Space | 🇸🇦 SA | apache-2.0 | 189 | 2025-06-10 | |
Arabic OCR Qwen2.5 VL 7B Vision | loay | 🌍 INTL | apache-2.0 | 129 | 2025-07-18 | |
arabic-small-nougat | MohamedRashad | 🌍 INTL | gpl-3.0 | 121 | 2024-11-28 |
Showing 20 of 66 · see all 66 on the interactive map · full table
Name | Org | Country | License | ⬇ Downloads | Updated | Links |
Arabic SBERT 100K | akhooli | 🌍 INTL | unknown | 16K | 2024-07-27 | |
GATE-AraBert-v1 | Omartificial-Intelligence-Space | 🇸🇦 SA | apache-2.0 | 14K | 2025-09-07 | |
asafaya/bert-base-arabic | asafaya | 🌍 INTL | unknown | 13K | 2023-03-17 | |
Arabic-STS-Matryoshka-V2 | Omar Elshehy | 🇪🇬 EG | unknown | 5K | 2024-12-29 | |
Arabic-Triplet-Matryoshka-V2 | Omartificial-Intelligence-Space | 🇸🇦 SA | apache-2.0 | 5K | 2025-09-07 | |
Arabic-Retrieval-v1.0 | Omar Elshehy | 🇪🇬 EG | apache-2.0 | 4K | 2024-12-19 | |
ARA-Reranker-V1 | Omartificial-Intelligence-Space | 🇸🇦 SA | apache-2.0 | 4K | 2025-04-03 | |
Mizan Rerank V2 | ALJIACHI | 🌍 INTL | apache-2.0 | 2K | 2026-10-10 | |
Arabic-all-nli-triplet-Matryoshka | Omartificial-Intelligence-Space | 🇸🇦 SA | apache-2.0 | 1K | 2025-01-23 | |
arabic english bge m3 | sayed0am | 🌍 INTL | mit | 1K | 2025-10-25 | |
silma-embedding-matryoshka-v0.1 | SILMA AI | 🌍 INTL | apache-2.0 | 1K | 2025-05-15 | |
Badr embedding v0 | somayaeltanbouly | 🌍 INTL | unknown | 1K | 2026-09-14 | |
Arabic-MiniLM-L12-v2-all-nli-triplet | Omartificial-Intelligence-Space | 🇸🇦 SA | apache-2.0 | 966 | 2025-06-10 | |
Arabert all nli triplet Matryoshka | Omartificial-Intelligence-Space | 🇸🇦 SA | apache-2.0 | 903 | 2025-01-23 | |
Marbert all nli triplet Matryoshka | Omartificial-Intelligence-Space | 🇸🇦 SA | apache-2.0 | 899 | 2025-01-10 | |
bojji | NAMAA-Space | 🇸🇦 SA | mit | 741 | 2025-06-16 | |
Muffakir Embedding V2 | mohamed2811 | 🌍 INTL | unknown | 619 | 2025-05-24 | |
Arabic mpnet base all nli triplet | Omartificial-Intelligence-Space | 🇸🇦 SA | apache-2.0 | 530 | 2025-01-23 | |
GATE-Reranker-V1 | NAMAA-Space | 🇸🇦 SA | apache-2.0 | 513 | 2025-04-03 | |
Arabic-labse-Matryoshka | Omartificial-Intelligence-Space | 🇸🇦 SA | apache-2.0 | 471 | 2025-01-10 |
Showing 20 of 1163 · see all 1163 on the interactive map · full table
Name | Org | Country | Size | License | ⬇ | Updated | Links |
prophet mosque library | ieasybooks | 🌍 INTL | 10K–100K rows | mit | 290K | 2025-05-14 | |
Waqfeya Library | ieasybooks | 🌍 INTL | 10K-100K books | mit | 96K | 2025-05-14 | |
shamela waqfeya library | ieasybooks | 🌍 INTL | 1K–10K rows | mit | 92K | 2025-05-14 | |
SARD | riotu-lab | 🇸🇦 SA | — | cc-by-nc-nd-4.0 | 46K | 2026-05-20 | |
Shamela4 Full DB | AuthenticIlm | 🌍 INTL | 10M–100M rows | mit | 40K | 2026-05-19 | |
wikiann | Rensselaer Polytechnic Institute | 🌍 INTL | 185,000 tokens | unknown | 36K | 2024-02-22 |
|
XNLI | Facebook AI Research | 🌍 INTL | 7,500 sentences | cc-by-nc-4.0 | 36K | 2024-01-05 |
|
Arabic Books | MohamedRashad | 🌍 INTL | 8.5k books | gpl-3.0 | 33K | 2024-11-28 | |
Aya Dataset | Cohere For AI Community | 🌍 INTL | 14,250 sentences | apache-2.0 | 23K | 2025-04-15 |
|
Quranic Recitation Data | zaibihassan | 🌍 INTL | 100K-1M clips | apache-2.0 | 15K | 2026-10-16 | |
Global-MMLU Lite | Cohere For AI | 🌍 INTL | 685 sentences | apache-2.0 | 14K | 2026-06-08 |
|
MMMLU | OpenAI | 🌍 INTL | 14,000 sentences | mit | 12K | 2024-10-16 | |
coda llm data | mohameddalii | 🌍 INTL | 1K–10K rows | apache-2.0 | 12K | 2026-10-12 | |
XStoryCloze | MetaAI | 🌍 INTL | 1,870 sentences | cc-by-sa-4.0 | 11K | 2025-07-23 |
|
TYDIQA | Google Research | 🌍 INTL | 25,893 sentences | apache-2.0 | 10K | 2024-08-08 |
|
MIRACL | unknown | 🌍 INTL | 2,061,414 documents | apache-2.0 | 8K | 2023-01-05 | |
karsl 502 arabic sign language v2 | FatimahEmadEldin | 🌍 INTL | — | unknown | 7K | 2026-01-20 | |
Dialectal Arabic Lahgtna v2 | oddadmix | 🇪🇬 EG | 3000h | unknown | 7K | 2026-08-04 | |
waqfeya library compressed | ieasybooks | 🌍 INTL | 10K–100K rows | mit | 7K | 2025-04-25 | |
prophet mosque library compressed | ieasybooks | 🌍 INTL | 10K–100K rows | mit | 6K | 2025-05-08 |
Showing 20 of 313 · see all 313 on the interactive map · full table
Name | Org | Country | License | Links | Notes |
voxlect-arabic-dialect-whisper-large-v3 | tiantiaf (USC) | 🌍 INTL | openrail | Whisper-large-v3 Arabic dialect classifier from the Voxlect benchmark. | |
hassaniya-punctuation-restoration | Emin009 | 🌍 INTL | apache-2.0 | Open-source punctuation-restoration model for Hassaniya Arabic. | |
flair-arabic-dialects-codeswitch-egy-lev | megantosh | 🌍 INTL | apache-2.0 | Flair and fastText part-of-speech tagger for Egyptian and Levantine code-switched Arabic. | |
OSMAN readability metric | Lancaster University | 🌍 INTL | unknown |
| Arabic readability metric analogous to Flesch, with parallel Arabic-English corpus. |
Agentic-AI-Design-Patterns | Muhannad-Khaled | 🌍 INTL | mit | The 21 agentic AI design patterns from Antonio Gulli's book, explained in Egyptian Arabic — with framework-free, runnable Python examples. | |
Ai-Quran-Video-Composer | WalidZein | 🌍 INTL | unknown | What if you can generate a a video with cinematic backgrounds behind the majestic verses of the Quran. | |
ai-rtl-resolver | miladniroee | 🌍 INTL | mit |
| Ai Chatbot RTL Resolver |
Al Kindi AI | Alpha to Digits | 🇦🇪 AE | proprietary | Arabic social listening intelligence product built on the Arabic.AI LLM. | |
AL-Khatma | oaokm | 🌍 INTL | mit | A library Specialized About Islamic | |
Al-Munir | zedsalim | 🌍 INTL | unknown |
| المنير: للاستماع وقراءة القرآن الكريم مع التفسير لتسهيل الحفظ والفهم والمراجعة |
al_quran_v3 | IsmailHosenIsmailJames | 🌍 INTL | other | A Flutter application for reading the Holy Quran, tracking prayer times, and managing Islamic practices. | |
albitaqat_quran | rn0x | 🌍 INTL | unknown |
| Complete data for 114 Quran surahs (Al-Bitaqat). |
Alfanous | Alfanous team | 🌍 INTL | lgpl-3.0 | Arabic search engine API for the Quran with simple and advanced queries. | |
algeria_69_wilayas | Mohamed-gp | 🌍 INTL | mit |
| 🇩🇿 All 69 Algerian wilayas + 1,541 communes as free JSON — official numbering, Arabic names, coordinates, dairas. |
Alkhalil Morpho Sys | unknown | 🌍 INTL | unknown | A morphosyntactic parser for Arabic words that can process both vocalized and non-vocalized texts | |
all-words-in-all-languages | eymenefealtun | 🌍 INTL | unknown | This repository contains all the words from every language that exists in the universe. | |
ALLAM based Retrieval Augmented Generation Arabic Conversational Alzheimer Assistance | ShahadAljohani | 🌍 INTL | cc-by-4.0 | ALLAM-RAG: Saudi Arabic conversational RAG system combining retrieval with the ALLaM LLM to support Alzheimer's patients and caregivers. | |
alpinejs-i18n | rehhouari | 🌍 INTL | mit |
| Easy i18n (Internationalization) for Alpine.js! |
alquranalkareem | alheekmahlib | 🌍 INTL | unknown |
| التطبيق الأمثل لقراءة القرآن الكريم |
Alyahmor | linuxscout | 🇩🇿 DZ | gpl-3.0 | Arabic flexional morphology generator; maintainer is Algerian. |
Showing 20 of 165 · see all 165 on the interactive map · full table
Name | Org | Country | License | Links | Notes |
Global-MMLU | Cohere For AI | 🌍 INTL | apache-2.0 |
| Multilingual MMLU-style benchmark of 42 languages (Arabic included) with culturally sensitive and culturally agnostic subsets. |
FLORES-101 | Facebook AI Research | 🌍 INTL | cc-by-sa-4.0 |
| Low-resource machine translation benchmark of 101 languages, with Arabic as one of them. |
xquad | HiTZ Center | 🌍 INTL | cc-by-sa-4.0 |
| Cross-lingual QA benchmark of 240 SQuAD paragraphs and 1190 question-answer pairs translated into ten languages including Arabic. |
ArabicMMLU | MBZUAI | 🇦🇪 AE | cc-by-nc-4.0 | Multi-task language understanding from school exams | |
MMCQA-SemEval27 | QCRI | 🇶🇦 QA | cc-by-nc-4.0 | This is the dataset for MMCultureQA, the SemEval 2027 shared task on culturally grounded visual question answering in English and Arabic. | |
Belebele-Fleurs | University of Würzburg | 🌍 INTL | cc-by-sa-4.0 |
| Belebele-Fleurs extends BeleBele into a spoken SLU benchmark with multilingual multiple-choice listening comprehension QA from speech. |
TunisianMMLU | LINAGORA | 🌍 INTL | cc-by-nc-sa-4.0 | MMLU translated into Tunisian Derja, LINAGORA (France); evaluated with lighteval. | |
AlGhafa Arabic LLM Benchmark (Native) | Open Arabic LLM Leaderboard (OALL) | 🌍 INTL | unknown | Native-Arabic multiple-choice evaluation tasks used as OALL leaderboard tasks. | |
Human-Translated Arabic MMLU | MBZUAI | 🇦🇪 AE | unknown | Human-translated MMLU into Arabic from MBZUAI, companion to ArabicMMLU. | |
MIRACL-VISION | NVIDIA | 🌍 INTL | cc-by-sa-4.0 |
| Multilingual visual document retrieval benchmark extending MIRACL with Wikipedia page images for 18 languages. |
Arabic EXAMS (OALL) | Open Arabic LLM Leaderboard (OALL) | 🌍 INTL | unknown | Arabic subset of the EXAMS multilingual school-exam benchmark, an OALL task. | |
ArAD | SpeechAntiSpoofingBenchmarks | 🌍 INTL | odc-by | Benchmark-ready packaging of the test split of the Arabic Audio Deepfake (ArAD) dataset: binary anti-spoofing on Arabic (primarily Levantine dialect) speech. | |
jev ar bench | atmaneayoub | 🌍 INTL | cc-by-4.0 | Arabic intent-routing benchmarks for MSA, Emirati, Saudi and code-switched Arabic | |
ASR Code Switch | Perle-ai | 🌍 INTL | mit | A curated benchmark of 1,200 code-switching utterances (300 per language pair) for evaluating commercial ASR systems on multilingual speech. | |
DarijaMMLU | MBZUAI-Paris Lab | 🇦🇪 AE | mit |
| MMLU translated into Moroccan Darija, 22k+ multiple-choice questions. |
fa en ar handwritten ocr v1 | saeid1999 | 🌍 INTL | cc-by-4.0 | A large, clean, augmentation-rich synthetic handwriting dataset for training and benchmarking OCR / HTR models on Persian (fa), Arabic (ar) and English (en). | |
OALL Arabic MMLU | Open Arabic LLM Leaderboard (OALL) | 🌍 INTL | unknown | GPT-translated Arabic MMLU copy from FreedomIntelligence, used as an OALL v1 task. | |
Quranic ASR Benchmark | Quran Lab | 🌍 INTL | other | Small benchmark set for evaluating ASR models on Quran recitation. | |
ALRAGE | Open Arabic LLM Leaderboard (OALL) | 🌍 INTL | unknown | Arabic retrieval-augmented generation evaluation set used in OALL v2. | |
EgyMMLU | UBC-NLP | 🌍 INTL | unknown | MMLU translated into Egyptian Arabic from UBC (Canada), released with NileChat. |
Showing 20 of 732 · see all 732 on the interactive map · full table
Title | Venue | Year | Topic | Links |
Systematic Review and Meta-Analysis of 2024–2025 Studies on AI Arabic Translation, Linguistics and Pedagogy | Frontiers in Computer Science and Artificial Intelligence | 2026 | survey, translation | |
A Holistic Assessment of the Carbon Footprint of Noor, a Very Large Arabic Language Model | BIGSCIENCE | 2026 | pretraining, evaluation | |
A Comparative Study of Pretrained Transformer Models for Quranic ASR: Speech Representations, Label Formats, and Dataset Composition | arXiv 2026 | 2026 | asr, diacritization, pretraining, speech, quran | |
A Corpus-Aligned Uthmani-to-Standard Quranic Word Mapping and a Deterministic Recitation Validator | arXiv 2026 | 2026 | asr, evaluation, quran | |
A Human-in-the-Loop Label Error Detection Framework Applied to Arabic-Script HTR Datasets | arXiv 2026 | 2026 | asr, ocr, evaluation | |
Abjad-Kids: An Arabic Speech Classification Dataset for Primary Education | arXiv 2026 | 2026 | evaluation, speech | |
ADAB: Arabic Dataset for Automated Politeness Benchmarking -- A Large-Scale Resource for Computational Sociopragmatics | arXiv 2026 | 2026 | evaluation | |
AHA-Memes: A Fine-Grained Multimodal Benchmark for Understanding Hate in Arabic Memes | arXiv 2026 | 2026 | evaluation, vision | |
Alexandria: A Multi-Domain Dialectal Arabic Machine Translation Dataset for Culturally Inclusive and Linguistically Diverse LLMs | arXiv 2026 | 2026 | chat, translation, evaluation, vision |
|
AlexandriaX 2026: The First Shared Task on Dialectal Arabic Machine Translation | arXiv 2026 | 2026 | chat, embedding, translation, evaluation |
|
Almieyar-Oryx-BloomBench: A Bilingual Multimodal Benchmark for Cognitively Informed Evaluation of Vision-Language Models | arXiv 2026 | 2026 | evaluation, vision |
|
Almieyar: A Culturally Grounded Benchmark for Multi-Dialect Arabic Speech Recognition | arXiv 2026 | 2026 | asr, evaluation, speech, vision | |
An End-to-End Hybrid Framework for Rumour Detection in Low-Resources Algerian Dialect | arXiv 2026 | 2026 | embedding, pretraining, evaluation | |
ArabDiscrim: A Decade-Long Arabic Facebook Corpus on Racism and Discrimination | arXiv 2026 | 2026 | vision | |
Arabic Prompts with English Tools: A Benchmark | arXiv 2026 | 2026 | pretraining, evaluation | |
ArabicDialectHub: A Cross-Dialectal Arabic Learning Resource and Platform | arXiv 2026 | 2026 | translation |
|
ArabicDialectSafety: A Dialect-Aware Benchmark for Arabic Content Safety Classification | arXiv 2026 | 2026 | safety, dialects, classification | |
AraDetox: A Multi-Dialect Arabic Detoxification Dataset | arXiv 2026 | 2026 | sentiment, evaluation |
|
ARAFA: An LLM-Generated Arabic Fact-Checking Dataset | arXiv 2026 | 2026 | evaluation | |
AraGenre 2026: A Hierarchical Definition-Guided Arabic Genre Classification Shared Task | arXiv 2026 | 2026 | evaluation |
Showing 20 of 125 · see all 125 on the interactive map · full table
Name | Country | Focus | Links |
King Saud University | 🇸🇦 SA | SaudiBERT, Saudi dialect corpora (STMC, SFC) | |
AIDA Lab (PSU) | 🇸🇦 SA | AI and Data Analytics lab at Prince Sultan University, Riyadh. | |
AIM Lab (NU-Q) | 🇶🇦 QA | AI and media research lab at Northwestern University in Qatar. | |
Ain Shams University | 🇪🇬 EG | Arabic NLP, sentiment analysis, NER research | |
AlooChat | 🇴🇲 OM | AI customer support chat agents for WhatsApp, Instagram, web and email, found via an Arabic conversational AI search in Oman. | |
AlQari | 🇸🇦 SA | Arabic-first document intelligence turning Arabic and English documents into structured, validated data. | |
Applied Innovation Center (AIC) | 🇪🇬 EG | Egyptian MCIT AI center; builds the Karnak LLM family and publishes models on Hugging Face (Applied-Innovation-Center). | |
Arabic.AI (Tarjama) | 🇦🇪 AE | Arabic-first autonomous AI - Pronoia Arabic LLM, Agentic AI platform | |
Arabot | 🇦🇪 AE | Conversational AI for Arabic - Arabic NLP chatbot engine | |
Aramco Digital | 🇸🇦 SA | Aramco digital arm, behind Metabrain generative AI assistant and AI partnerships. | |
ARBML | 🇸🇦 SA | Democratizing Arabic NLP - masader, klaam, tkseem | |
ASAS AI | 🌍 INTL | Saudi AI company offering Arabic models, automation and in-Kingdom infrastructure. | |
Astra Tech (Botim) | 🇦🇪 AE | Abu Dhabi tech group behind Botim, building AI assistant products. | |
AtlasIA | 🇲🇦 MA | Behind AL Atlas Moroccan Darija models | |
AUB MIND Lab | 🇱🇧 LB | Foundational Arabic NLP models - AraBERT, AraGPT2, AraELECTRA | |
Calfa | 🌍 INTL | OCR service extracting printed and handwritten text from scanned documents, including Arabic script. | |
CAMeL Lab (NYU Abu Dhabi) | 🇦🇪 AE | CAMeLBERT, camel_tools, morphological analysis | |
CAMeL Lab, NYUAD | 🇦🇪 AE | Arabic NLP tools and models - CAMeLBERT, camel_tools | |
Clusterlab AI | 🌍 INTL | Behind 101 Billion Arabic Words Dataset and InstAr-500k | |
Cohere | 🌍 INTL | Multilingual LLMs - Command R Arabic, RAG optimization |
Skills shipped in this repo:
Skill | What it does | Install |
arabic-ai-advisor | Use when choosing an Arabic LLM, ASR, TTS, OCR, embedding model, dataset or tool — recommends ranked options from the Arabic AI Atlas with licenses, dialect coverage and on-device fit. |
|
arabic-dialect-prompts | Use when writing system prompts or few-shot examples for an Arabic-speaking agent in a specific dialect, or when a model answers in formal Arabic (MSA) instead of the requested dialect. |
|
arabic-token-cost | Use when estimating or comparing how many tokens Arabic text costs across LLM tokenizers, or when Arabic prompts seem expensive, truncated or slow compared with English. |
|
rtl-bidi-lint | Use when reviewing or generating Arabic text, UI strings, HTML or Markdown that mixes RTL with LTR content, to catch digit mixing, unbalanced bidi isolates, stray LTR marks and hardcoded dir="ltr". |
|
tashkeel-check | Use when Arabic text may need diacritics (tashkeel) added or stripped, for TTS input, learner material, Quran or poetry, or when diacritization looks partial or inconsistent. |
|
Other Arabic agent skills in the wild:
Showing 20 of 35 · see all 35 on the interactive map · full table
Name | Org | Notes | Links |
ArabAgentSkills | ArabAgentSkills | Public source-backed Arab-world agent skills covering APIs, marketing, sales, support and localization. | |
ArabGuard | d12o6aa | Python SDK protecting LLMs and chatbots from prompt injection in Arabic text. | |
Arabic Content Studio | smeseik-ai | Claude Skills for Arabic content teams covering production, culture, SEO and brand voice. | |
Arabic Dict MCP | arnizamani | MCP server for grounded Arabic dictionary lookup wrapping arramooz (MSA verbs and nouns). | |
Arabic Scholar MCP Server | EngDawood | MCP server for searching Arabic academic research, articles and dissertations across sources. | |
Arabic Video Subtitles Skill | EngDawood | Skill with Netflix-style Arabic subtitling guidelines and an SRT/VTT checker. | |
Arabic Word Production Skill | Bannovich | Deterministic agent skill and plugin for Arabic-first and bilingual Word DOCX production. | |
arabic-bidi-engineering | MosaabGalmod | Agent skill for correct Arabic RTL/BiDi in chat, terminal and documents | |
arabic-pii-py | Aajil Labs | Local-first reversible PII tokenization for Arabic and Gulf data before it reaches an LLM. | |
ArabiMaak MCP | deepdiver4ai | Gulf Arabic language MCP server for word, phrase and cultural-context lookup. | |
awesome-arabic-claude-skills | EngDawood | Curated library of open-source Arabic skills for Claude Code and agents | |
Azan MCP | Ahmed Eltaher | Lightweight MCP library for Islamic prayer times and Qibla for AI agents; not Arabic-language specific. | |
Claude Adhkar | nosseralaa7-rgb | Shows Arabic-script adhkar in the Claude Code spinner with a terminal font that renders Arabic. | |
Claude Arabic Writing | Ahmed Dabak | Claude Agent Skill for natural, grammatically correct Arabic writing and translation. | |
Claude Code RTL Extension | yechielby | VS Code and Cursor extension adding RTL support for Hebrew and Arabic in Claude Code. | |
Claude Desktop RTL Patch (macOS) | toboly | Adds auto-detected Hebrew and Arabic RTL support to Claude Desktop on macOS. | |
Dorar Hadith MCP | ibnsaleem29 | Claude extension and MCP server for Hadith research including isnad and takhrij via Dorar. | |
Fanar MCP Server | danijeun | MCP server exposing Fanar API tools such as Islamic RAG and image generation. | |
Hadith MCP (ovehbe) | ovehbe | MCP server for searchable, citation-safe hadith text. | |
Hurmoz | Moshe-ship | Collection of 63 Arabic skills for the Hermes Agent framework. |
Gaps nobody has filled yet. Add an entry that matches a rule and it moves to Recently filled. Rules live in data/wanted.yaml.
Gap | Why | Rule |
No openly licensed text-to-speech model with a Kuwaiti or Qatari voice is listed. |
| |
The only Moroccan OCR dataset listed has no clear license, so no openly licensed one is available. |
| |
A Sudanese ASR model exists, but no Sudanese speech dataset to train or test one is listed. |
| |
Mauritania has no listed model, dataset, or tool of any kind. |
| |
No openly licensed Yemeni dataset for speech recognition or synthesis is listed. |
| |
No Arabic embedding model targets the Sudanese, Yemeni, or Iraqi dialects. |
| |
The only TTS covering Iraqi has no clear license, so no openly licensed one is available. |
| |
No dataset from or about Libya is listed. |
| |
No openly licensed dataset of Syrian Levantine Arabic is listed. |
| |
No speech recognition model from Palestine is listed. |
| |
Omani datasets are listed, but none has a clear open license. |
| |
The only Kuwaiti dataset listed has no clear license. |
| |
The only Bahraini dataset listed is non-commercial, so no openly licensed one is available. |
| |
Nothing from Comoros, Djibouti, or Somalia is listed. |
| |
No openly licensed benchmark covers Arabic handwriting recognition. |
|
385 models with known lineage · explore the tree · add a base model with the base_model field (CONTRIBUTING)
Base family | Models | Most downloaded |
Qwen · كوين | 65 | qwen3-asr-arabic-uae (42K ⬇) |
BERT · بيرت | 58 | AraBERTv02 (558K ⬇) |
Whisper · ويسبر | 56 | whisper-large-v3-turbo (6.3M ⬇) |
Gemma · جيما | 24 | gemma4 e4b claims comparison (342K ⬇) |
From scratch · من الصفر | 19 | aragpt2-base (13K ⬇) |
Llama · لاما | 15 | Llama 3.3 (257K ⬇) |
Cohere | 9 | cohere-transcribe-arabic-07-2026 (100K ⬇) |
T5 · تي 5 | 9 | AraT5 (2K ⬇) |
XTTS | 7 | XTTS-v2 (6.7M ⬇) |
BGE | 4 | arabic english bge m3 (1K ⬇) |
F5-TTS | 4 | f5tts-algerian-darja (12K ⬇) |
XLS-R · إكس إل إس-آر | 4 | |
NLLB | 3 | darija to english 2 (21 ⬇) |
SeamlessM4T | 2 | SeamlessM4T v2 (305K ⬇) |
wav2vec · واف تو فيك | 2 | wav2vec2 quran phonetics (11K ⬇) |
BLOOM · بلوم | 1 | darija text generation (25 ⬇) |
E5 | 1 | algerianME5 (68 ⬇) |
Falcon · فالكون | 1 | Falcon Arabic (11K ⬇) |
Mistral · ميسترال | 1 | Mistral 7B v0.1 arabic (22 ⬇) |
MMS · إم إم إس | 1 | |
Other · أخرى | 99 | Multilingual Chatterbox (1.7M ⬇) |
Edit the YAML in data/.
Run uv run python scripts/build.py build (it uses the committed Hugging Face cache; the nightly job refreshes metrics).
Open a PR.
Never hand-edit README.md, docs/tables/, assets/, dist/ or data/.cache/: the build and the nightly job write them, and CI rejects drift.
Code: MIT. Data: CC BY 4.0.
Awesome Arabic NLP, the curated reading list this atlas grew alongside.
Generated 2026-10-17 from 3302 entries. Do not edit README.md by hand.
Fetch one full Arabic AI Atlas entry by its id (e.g. "jais-30b"), including links and metrics.
Returns {"error": "unknown id", "id": ...} when no entry has that id. Use search to find ids.
Example: get(id="jais-30b")
| Name | Required | Description | Default |
|---|---|---|---|
| id | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries the full behavioral burden. It discloses the rich returned payload ('links and metrics') and the exact error shape ({"error": "unknown id", "id": ...}) for missing ids, but does not mention auth requirements, rate limits, or read-only nature explicitly.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Three tight sentences: purpose, error behavior, and routing, followed by a short example. Front-loaded with the core capability and zero filler.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a single-param read tool with no output schema or annotations, the description supplies purpose, error semantics, sibling routing, and an example. It stops short of defining auth or rate-limit expectations, but it is sufficient to call the tool correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 0% and the sole parameter `id` is undocumented in the schema; the description compensates by specifying the id format and giving a concrete example ('jais-30b'). Baseline for a one-param tool is 4, and the example lifts it, though no further constraints (length, allowed charset) are given.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
States a specific verb (fetch) and resource (one full Arabic AI Atlas entry) with scope ('by its id'). Names a concrete id example ('jais-30b'), making the tool's domain unambiguous against the sibling tools search and recommend.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Explicitly routes the agent to `search` to find ids, which is a clear alternative-tool pointer. It lacks an explicit when-not statement (e.g., 'do not use for listing or bulk retrieval'), but the conditional routing is strong.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recommend entries from the Arabic AI Atlas for a task, ranked with an explanation.
Score = 3 if the task is in the entry's tasks, +2 if the dialect matches, +1 if the task
word appears in its notes; zero-score entries are dropped. Each result carries score
and why. Ties go to models (llm, asr, tts, ocr, embedding) over datasets, benchmarks,
tools and orgs, then to downloads.
task: e.g. chat, tts, asr, ocr, embedding, translation. dialect: msa, egy, gulf, lev, magh, iraqi, sudanese, yemeni, classical, mixed (optional; entries lacking dialect data still match on task). on_device: true keeps only entries marked on-device (phone or laptop CPU); false drops those; omit for no filter. license_filter: "open" excludes proprietary/unknown licenses; any other string must equal the license exactly (optional). type: exact entry type, e.g. "tts" to get models only, "dataset" for training data (optional).
Example: recommend(task="tts", type="tts", on_device=true, license_filter="open")
| Name | Required | Description | Default |
|---|---|---|---|
| task | Yes | ||
| type | No | ||
| limit | No | ||
| dialect | No | ||
| on_device | No | ||
| license_filter | No |
| Name | Required | Description |
|---|---|---|
| result | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations present, the description carries the full burden and does so well: it discloses the exact scoring algorithm, that zero-score entries are dropped, the tie-break preference order (models over datasets/benchmarks/tools/orgs, then downloads), and that results carry `score` and `why`. This is unusually rich behavioral disclosure for a tool with zero annotation coverage.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Front-loaded with the purpose and scoring rule, then grouped parameter notes, then a concrete example. The scoring and tie-break sentences earn their place, though the parameter list is a bit dense and could be tightened.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a 6-parameter filtering+ranking tool with an output schema, the description covers the ranking logic, filter semantics, and return fields (`score`, `why`), so an agent has almost everything needed. The undocumented `limit` default of 3 is the one remaining gap.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 0%, so the description must compensate, and it documents five of six parameters with real semantic detail (allowed dialect values, the three-state on_device semantics, license_filter exact-match behavior, type as an exact entry type). It omits the `limit` parameter entirely, leaving one gap.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
States a specific verb and resource ('Recommend entries from the Arabic AI Atlas') plus the ranking behavior, which is enough to separate it from the sibling tools search and get, which are lookup tools. An agent can tell this is a ranked-recommendation tool without opening the schema.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description gives concrete guidance on how to drive the tool: the scoring formula (task +3, dialect +2, notes +1), tie-breaking rules, and per-parameter usage including a worked example. It does not explicitly state when to prefer this over the siblings search/get, so it falls short of full when/when-not routing.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Search the Arabic AI Atlas, a curated catalogue of Arabic models, datasets, benchmarks, tools and organizations with Hugging Face download metrics.
Case-insensitive substring match over name, org, notes, tasks and tags; the optional type/country/modality filters are exact. Results are ordered by downloads, most first.
Valid values: type: llm, asr, tts, ocr, embedding, dataset, benchmark, tool, agent-skill, org country: SA, AE, EG, QA, MA, JO, TN, LB, KW, OM, BH, INTL modality: text, speech, vision, multimodal, none
Example: search(query="speech", type="asr", limit=5)
| Name | Required | Description | Default |
|---|---|---|---|
| type | No | ||
| limit | No | ||
| query | Yes | ||
| country | No | ||
| modality | No |
| Name | Required | Description |
|---|---|---|
| result | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden and does well: it discloses match semantics (case-insensitive substring on name/org/notes/tasks/tags), that type/country/modality are exact filters, and the result ordering (downloads, most first). It omits any note on auth, rate limits, or what 'limit' can/cannot do, which keeps it below 5.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Front-loads the purpose, then match/ordering semantics, then the enum reference block, then a concrete example. Every sentence adds information an agent needs; nothing is boilerplate.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
An output schema exists, so return-value explanation is correctly omitted, and the description covers filtering, matching and ordering. The only gap is pagination/limit behavior, which is minor for a catalogue search but not nothing.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 0%, so the description must compensate and largely does: it explains that 'query' is a substring match over specific fields, names the three filter params as exact matches, and enumerates every valid value for type, country and modality. Only 'limit' is left without explicit semantics (default 10 is visible in the schema but no max/behavior note).
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
States a specific verb (search) and resource (Arabic AI Atlas catalogue), and enumerates the entity kinds indexed (models, datasets, benchmarks, tools, organizations). Clear what it does, but it never explicitly distinguishes itself from the sibling tools 'recommend' and 'get'.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
Usage is implied by the match semantics and the worked example, but there is no explicit when-to-use / when-not-to-use guidance and no mention of when to prefer 'recommend' or 'get' instead. An agent can infer the use case but must do the routing itself.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.0getrecommendsearchScored across 3 tools
The three tools have clearly distinct purposes: `search` does substring matching with exact filters, `recommend` does task/dialect ranked scoring with explanations, and `get` fetches a single entry by id. The descriptions even cross-reference each other (get says 'use search to find ids'), leaving no realistic selection ambiguity.
All three names are single lowercase verbs (search, recommend, get) in a uniform style with no mixing of conventions. The pattern is predictable and readable.
Three tools is on the thin side, but for a read-only catalogue the discovery/ranking/retrieval split earns each tool's place. Nothing redundant or missing at the surface level.
The read-only domain is well covered: broad search with filters, semantic-ish task recommendations, and full-detail retrieval. Minor gaps exist (no way to list all entries, browse by organization, or enumerate valid filter values beyond the examples), but agents can work around these via search.
Find the right MCPs, tools and skills for any AI task - discovery over a 14k+ resource catalog.
Search a curated directory of agent tools, MCP servers, APIs, protocols and skills. No auth.
Discovery for the agent ecosystem: findable businesses, tools, knowledge, agent spaces, skills.
Search, inspect, recommend, and explain rated AI tools through Agent Radar.