Jina AI
Also known as: Jina, Jina Search Foundation
Search foundation models and APIs. Reader turns URLs into markdown ready for an LLM, and Jina adds web search, multimodal embeddings and rerankers, served by API and MCP or deployed air gapped.
Jina AI builds search foundation models and serves them through a single API key. Reader turns any URL or uploaded PDF or HTML file into clean markdown for a model, loading the page in a browser engine with controls for waiting on elements, cookies, viewport and custom scripts. A companion endpoint returns web search results in the same form. Its embedding models are multimodal and multilingual, its rerankers reorder retrieved results by relevance, and the API also offers classification, batch embedding jobs and an OpenAI compatible chat completions endpoint.
Developers reach Jina through its REST API, an MCP server at mcp.jina.ai, a CLI and llms.txt guides for coding agents. Every model has open weights on Hugging Face and can be deployed fully air gapped in private infrastructure. The models can also run in a customer's own AWS or Azure account, and the hosted API can keep a request's processing inside the EU.
Jina AI was acquired by Elastic in October 2025. Pricing is based on tokens. API keys come with free tokens, tokens are topped up as you go, and paid and premium keys raise rate limits.
Vendor details
Canonical URL
https://jina.ai
Category
Agent infrastructure
Subcategory
Search and retrieval
Funding status
Acquired by Elastic in October 2025, according to Jina's own model guide; jina.ai now carries Elastic's copyright.
Company status
acquired
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
One API key covers Reader (r.jina.ai), web search (s.jina.ai), embeddings, reranking, classification and batch jobs, with a published OpenAPI specification, an MCP server at mcp.jina.ai, a CLI and llms.txt guides. Models can be deployed in the customer's AWS or Azure account and fully air gapped in private infrastructure.
In practice
Your agent needs to read a web page as clean text. You prepend the Reader host to any URL and get model ready markdown with JS rendered server side, ready to drop into a prompt.
Your RAG results are noisy. You retrieve wide with Jina embeddings, then apply the reranker to rescore and shrink to a precise set before handing context to the generator.
You want one retrieval layer across an automation flow. Jina's single key covers Reader, embeddings and reranking with shared tokens, and your own code chains them into a pipeline from URL to answer.
Sources & related URLs
Agentic Index coverage score
5.0 / 14 capabilities · 36%
| Integrations & Tool Calling | Partial |
|---|---|
|
An agent gets read access to the web and to Jina's own models as tools. Reader turns a URL into markdown, s.jina.ai returns search results, and the embedding, reranking and classification APIs are reachable directly or through the MCP server at mcp.jina.ai. The tools only retrieve. They take no actions in other systems and have no per user credentials to scope. Sourcejina.airead 2026-09-22 |
|
| Workflow Orchestration | Not documented |
|
Each call to Jina's API stands alone. Model calls for embeddings, rerank, classify, batch jobs and chat completions run independently, as do Reader requests. The API has no way to sequence, branch or route steps into a workflow, so pipelines that chain Reader, embeddings and reranking are built in the customer's own code or tools. Sourceapi.jina.ai/openapi.jsonread 2026-09-22 |
|
| Knowledge Grounding & RAG | Partial |
|
Customers can build a retrieval layer over their own documents from Jina's models. These include multimodal, multilingual embeddings, rerankers, a classifier and a segmenter for chunking, with batch embedding jobs and fine tuning of classifiers on the customer's data. Jina keeps no index or store of the customer's corpus, and the vectors live in the customer's own database or Elasticsearch. Web search and Reader work over the web, not the customer's corpus. Sourcejina.ai/models/llms.txtread 2026-09-22 |
|
| Human Oversight & Guardrails | Not documented |
|
The APIs read, embed and rank content and take no actions in other systems. Jina names no approval step, review queue, escalation rule or policy constraint on what an agent does. A token budget can be set per request to control cost. Sourcejina.airead 2026-09-22 |
|
| Security, Identity & Governance | Partial |
|
Customers are isolated in dedicated trust zones, with encryption at rest and in transit using keys for each customer. Jina runs internal and external penetration testing, secure development training and design reviews, and its terms offer a GDPR data processing agreement. A per request option keeps processing inside the EU. Access is by API key, and Jina names no SSO, roles or attestation. Sourcejina.ai/legalread 2026-09-22 |
|
| Observability & Auditability | Not documented |
|
The API dashboard shows a key's available tokens and usage, and rate limits are published by tier. Token balances and rate limits are billing views. Jina names no per request log, trace or audit record of what an agent sent and received, and no way to inspect prompts, tool calls and outputs step by step. Sourcejina.ai/api-dashboard/pricingread 2026-09-22 |
|
| Memory & State Persistence | Not documented |
|
No agent context carries between calls to Jina's APIs. Stored classifiers and batch job outputs are model artifacts and results, and the API has no memory layer for agents. Sourceapi.jina.ai/openapi.jsonread 2026-09-22 |
|
| Deployment & Data Residency | Full |
|
Models run where the customer chooses. By Jina's account, every model can be deployed fully air gapped in private infrastructure with no external network dependency. Customers can also deploy through their own AWS or Azure account, billed through the cloud provider, and the hosted API has a per request option that keeps infrastructure and data processing inside the EU. Sourcejina.ai/models/llms.txtread 2026-09-22 |
|
| Prebuilt Agents, Templates & Packs | Not documented |
|
Models (embeddings, rerankers, readers, OCR) and a guide to choosing among them are what Jina publishes. Customers build with these components, and there are no ready made workflows, templates or role specific agents. Sourcejina.ai/models/llms.txtread 2026-09-22 |
|
| Triggers & Channel Coverage | Not documented |
|
Requests are answered as they arrive. The API has no schedule, webhook, event subscription or channel that starts an agent, so work begins only when a person initiates it with a request. Sourceapi.jina.ai/openapi.jsonread 2026-09-22 |
|
| Model Flexibility & Routing | Full |
|
Whatever model the customer's agent runs can take retrieval from Jina, through its API and an MCP server any client can add. Within Jina's API the customer picks among many embedding, reranking and reader models by setting the model parameter, and the API includes an OpenAI compatible chat completions endpoint. Sourcejina.ai/models/llms.txtread 2026-09-22 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
An OpenAPI specification is published for Jina's API, covering embeddings, rerank, classify, train, models, batch jobs and OpenAI compatible chat completions. The Reader and search endpoints sit at r.jina.ai and s.jina.ai, and Jina also offers an MCP server at mcp.jina.ai, a CLI and llms.txt guides for coding agents. Sourceapi.jina.ai/openapi.jsonread 2026-09-22 |
|
| Testing, Debugging & Optimization | Not documented |
|
Research papers and benchmark results cover Jina's own models, and a speed test covers Reader. None of these tests or scores the customer's agent, and Jina names no testing, scoring or quality gate for the customer's agent workflows. Sourcejina.ai/models/llms.txtread 2026-09-22 |
|
| Browser & Computer Use | Partial |
|
The Reader loads a page in a browser engine Jina runs, with options for viewport size, cookies for authenticated pages, waiting for a selector, removing elements and running custom JavaScript on the page before it extracts the content as markdown. It works as a headless browser fetch for extraction, and an agent cannot use it to click through or operate an interface step by step. Sourcejina.airead 2026-09-22 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
Jina AI released jina-ocr-v1, a 3.4 billion parameter vision language model that converts document pages to Markdown in a single pass and is built to run on low budget GPUs.
Bears on: Agent capability
View sourceJina AI released jina-reranker-v3.5, a 0.6-billion parameter listwise reranker featuring hybrid attention and self-distillation. The updated model reranks up to 1.56x faster than version 3 and improves performance on semi-structured retrieval tasks.
Bears on: Agent capability
View sourceJina AI introduced Jina On-Prem, a fully self-contained installation suite that packages all 28 of its embedding and reranking models into ready-to-deploy Docker containers. The containers run entirely offline with zero telemetry, license servers, or outbound connections. They serve models via API schemas compatible with OpenAI, Cohere, Voyage AI, Gemini, and the Elastic Inference Service.
Bears on: Deployment / data residency
View sourcePricing
Free API key tokens · pay-as-you-go token top-ups · premium keys for higher limits · AWS and Azure deployment
tokens processed
Included quota
Free tokens on a new API key; rate limits by tier, for example Reader at 20 requests a minute without a key, 500 with a free or paid key and 5,000 with a premium key.
What is public
The billing model and rate limits by tier are public; top-up bundle prices sit on a client-rendered dashboard.
Billing mechanics
One API key and one token balance cover every service; top up as you go, upgrade to a premium key for higher limits, or deploy models in your own cloud account.
Cost watchouts
One token balance is drawn down by every service, and each web search request costs a fixed block of tokens (from 10,000), so an agent that searches on every loop drains the balance quickly; set a per-request token budget.
Variable cost rationale
Cost scales with tokens processed across retrieval steps, but published per model rates such as $0.02 per million for embeddings keep unit economics predictable, and a shared token pool makes spend easy to cap.
Overage / add-ons
Usage draws down a shared token balance across all services; top ups replenish it, with rate limits set by tier.
Sales call required
Mixed (some tiers require a call)
Free / trial
Free trial with millions of tokens and no card, plus a keyless rate limited Reader tier
Lowest paid plan
Pay-as-you-go token top-ups on the API dashboard
Commercial notes
Acquired by Elastic in October 2025; jina.ai now carries Elastic's copyright.
Key ambiguities
Top up bundle prices sit on a client rendered dashboard, so the price per token is not established. Ask Jina for current top up pricing.
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…
- AlphaBitCore — AI control plane for regulated financial firms: one gateway enforces…
Alternatives to Jina AI
The closest documented capability profiles to Jina AI among agent infrastructure platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Tavily4.5 / 14Adds documented Observability & Auditability and Prebuilt Agents, Templates & Packs
- AgentOps5.0 / 14Adds documented Observability & Auditability and Testing, Debugging & Optimization
- Cognee7.0 / 14Adds documented Workflow Orchestration and Observability & Auditability, among others
- GPT Researcher7.0 / 14Adds documented Workflow Orchestration and Observability & Auditability, among others
- Hyperspell7.0 / 14Adds documented Observability & Auditability and Memory & State Persistence, among others
- Klavis AI7.0 / 14Adds documented Human Oversight & Guardrails and Triggers & Channel Coverage, among others
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded