How do you build Arabic AI with AI agents?

Key facts

Endpoint	POST /v1/chat/completions with tools; Arabic and mixed-language prompts use the same API
Multilingual models	30+ models in the catalogue; check each model entry for Arabic guidance
Embeddings	plugsky-embed-multilingual covers Arabic and mixed-language retrieval
Free tier	Free plan with 2 free AI models, no card required
Pricing model	Flat monthly self-serve plans with unlimited fair-use usage
Dialect policy	Prompt the target dialect explicitly and evaluate per dialect rather than assuming MSA transfers
Governance	Scoped keys, audit logs and region pinning for residency requirements
Roadmap	Audio and fine-tuning endpoints are coming soon

TL;DR

• Keep the standard agent loop; the Arabic work is prompts, normalization and evaluation.
• Choose Modern Standard Arabic or a named dialect deliberately — do not leave it to chance.
• Normalize Arabic input before classification: unify alef forms, strip diacritics and tatweel.
• Write tool descriptions bilingually so parameter meaning survives translation.
• Score quality per dialect and per script mix before launch.

How it works, step by step

1. Decide which Arabic the agent must serve: Modern Standard Arabic, one dialect, or a code-switching mix.
2. Write the system prompt in that Arabic variety and include one English mirror for maintainers.
3. Normalize incoming text: unify alef and ya forms, remove diacritics and tatweel, standardize digits and whitespace.
4. Define tools with bilingual descriptions and keep identifiers, enums and ids in Latin script.
5. Run the agent loop on a multilingual model, returning responses with the requested script and digit style.
6. Evaluate with Arabic-native test sets per dialect, including names, numbers, dates and mixed-script queries.

Original data

Try it yourself

Open the best model for agents selector →

What actually changes for Arabic agents

The orchestration pattern does not change; four language details do:

• Variety: Modern Standard Arabic is formal and broadly readable, while Gulf, Egyptian, Levantine and Maghrebi dialects differ in vocabulary and idiom. Pick one target and say so in the prompt.
• Morphology: Arabic packs meaning into prefixes and suffixes, so word-level keyword matching and naive token budgets behave differently than in English. Budget context in characters or tokens measured on your own corpus.
• Script handling: alef variants, ya variants, diacritics and tatweel create multiple surface forms for the same word. Normalize before any matching or routing decision.
• Direction: responses render right to left, and mixed Arabic-Latin strings reorder visually. Keep ids, codes and URLs Latin and isolate them in the UI.

Implementation details that matter

Practical choices for an Arabic-first agent:

• Write the system prompt in the target Arabic variety; add a short English mirror as a comment for maintainers.
• Describe each tool twice: an Arabic description for the model's reasoning and a Latin identifier for execution.
• Use plugsky-embed-multilingual for any semantic memory, routing examples or retrieval inside the agent.
• Normalize numbers and dates explicitly, and state which digit style the response should use.
• Keep enums and machine values in Latin; only user-facing strings should be Arabic.
• Test with real inputs from your users, not translated English samples — translated text hides dialect problems.

Evaluating Arabic agents

An aggregate score hides dialect failure, so segment the evaluation:

• Intent and tool selection: per dialect and per script mix (pure Arabic, Arabic with English terms, Arabizi).
• Response quality: fluency, formality match and correct honorifics for the target audience.
• Entity accuracy: names, places, dates and amounts, which are where normalization bugs surface.
• Structured output: JSON validity and field completeness when the agent calls tools.
• Refusal behavior: does the agent ask for clarification rather than guessing on ambiguous dialect input?

Freeze Arabic test sets alongside your English ones and re-run both on every prompt or model change.

Limitations

Multilingual models are not equally strong in every Arabic variety, and quality varies by model and task. Honest constraints:

• Dialect performance can differ sharply from Modern Standard Arabic; verify on your own data before promising coverage.
• Arabizi (Latin-script Arabic) may need explicit examples in the prompt or a normalization step.
• Arabic speech, translation-specific endpoints, audio and fine-tuning are coming soon; voice-first Arabic products need another stack today.
• Diacritic stripping improves matching but loses pronunciation detail, which matters for names and religious or legal text.
• Region pinning and scoped keys still apply; Arabic language support does not change residency obligations.

Honest comparison

Capability	Arabic agent on Plugsky	English-only stack plus translation	Single-model local deployment
Model choice	30+ multilingual models behind one API	Translation hop adds latency and errors	Only models you can host
Dialect policy	Explicit in prompts and evaluation	Flattened to one variety	Whatever the model learned
Embeddings	plugsky-embed-multilingual for Arabic retrieval	English embeddings on translated text	Depends on the model you serve
Normalization	Your pipeline, applied before routing	Usually ignored	Your responsibility
Governance	Scoped keys, audit logs, region pinning	Vendor-dependent	Full control, full ops burden

Frequently asked questions

Does Plugsky support Arabic prompts?

Yes. Arabic and mixed-language prompts go through the same OpenAI-compatible chat completions endpoint; quality depends on the model you choose and how you write the prompt.

Which model should I use for Arabic?

There is no single answer — evaluate two or three candidates from the catalogue on your own Arabic test set. Use plugsky-embed-multilingual regardless of which generation model you pick.

Should I use Modern Standard Arabic or a dialect?

Choose based on your users. Modern Standard Arabic is the safe default for formal or cross-market products; dialect-aware prompts are better when users write informally.

How do I handle Arabizi?

Add Arabizi examples to the prompt or evaluation set, or normalize Latin-script Arabic before routing. Test explicitly, because performance is model-dependent.

Do I need special tool schemas for Arabic?

Descriptions should be written for the model's language while identifiers and enums stay in Latin. This keeps execution reliable and reasoning natural.

Is there Arabic speech support?

Not yet — audio endpoints are coming soon. Today, Arabic products should handle text and integrate separate speech tooling if needed.

Can I prototype Arabic agents for free?

Yes. The free plan includes two free models with no card, and a 14-day full-access trial lets you compare multilingual models on your own Arabic data.