DevRel newsletter — September 2026

Hello from the Elastic DevRel team! In this newsletter, we cover jina-ocr-v1, the latest blogs and videos, and upcoming events like Elastic{ON}.

What’s new? #

• jina-ocr-v1 is now available as a service and on-prem with layout-aware document parsing for tables, math, handwriting, and 100+ languages in one model.
• The Elastic CLI gives you a single command for every public Elasticsearch, Kibana, and Cloud API with keychain credentials, schema validation, and policy controls built in.Agent Skills have been rebuilt on top of it.
• Precomputed Knowledge Indicators stored in Elasticsearch let agents answer questions using 93% fewer tokens and at a third of the latency, compared to standard retrieval augmented generation (RAG) over full documents.
• Large language model (LLM) spans in Elastic APM now surface prompts, responses, and per-call token counts directly in the trace waterfall.
• The Elastic Security Labs team instrumented 1,100 developer machines with Cursor hooks and collected 13 million tool call events. The post walks through what they found and how to replicate the setup with Elastic Agent.

jina-ocr-v1: One model for layout, tables, math, and 100+ languages

jina-ocr-v1 is now available in the Elastic Inference Service. It is an end-to-end document parser that takes scanned pages, document photos, presentation slides, and handwritten notes and returns structured Markdown in a single pass. No pipeline of separate specialist models needed.

What separates it from traditional OCR is how it handles page structure. Standard OCR reads characters left to right and often scrambles multi-column layouts, sidebars, and mixed-content pages. jina-ocr-v1 understands the page layout first and then reads in the order a human would. Tables come out as HTML. Mathematical formulas come out as LaTeX. Text from embedded images (book covers, logos, incidental photos) is recognized but left out of the output.

The model has 3.4 billion parameters total but uses a mixture-of-experts design, so only 570 million are active at inference time. It runs at the cost and speed of a 570M model while benefiting from the breadth trained across 3.4B parameters.

On olmOCR-bench, jina-ocr-v1 scores 83.4, which is the highest of any model with fewer than 600 million active parameters. On OmniDocBench, it scores 91.14,which is above GPT-5.2 at 86.59. The models that beat it on both benchmarks have significantly more active parameters.

Language support covers more than 100 languages, including scripts that frequently trip up single-language OCR like Arabic, Chinese, Japanese, Korean, Thai, Hindi, Greek, Cyrillic, Turkish, Czech, and more. Mixed-language pages are handled in one pass without extra configuration.

There are four ways to use it:

1. Elastic Inference Service: Available now via the chat_completions inference endpoint; native PDF support is coming soon
2. Jina API: Standard HTTP access with 10 million free tokens to get started
3. Local or on-prem: Download from Hugging Face under CC BY-NC 4.0, or use Jina On-Prem for commercial deployments
4. Jina AI Reader API: Add X-Respond-With: jina-ocr-v1 to the request header; this path converts PDFs and HTML pages to images automatically

Architecture details, benchmark tables, and code for each access method are in the release post.

Elastic CLI: One command for every Elasticsearch, Kibana, and Cloud API

The Elastic CLI is now in technical preview. One install gives you a single entry point for public Elasticsearch, Kibana, and Elastic Cloud API. Install it with npm install -g @elastic/cli.

The design goal was a consistent CLI. Every command works the same way: JSON in (via stdin or --input-file), JSON out (via --json), schema validation before any request leaves your machine, and a non-zero exit code on failure. Try elastic es search and you already know how elastic kb data-views list and elastic cloud serverless projects create behave.

A few things worth noting for anyone already thinking about agents:

• Credentials stay in the OS keychain: Running elastic config context add writes your API key to macOS Keychain, Linux Secret Service, or Windows Credential Manager and puts a $(keychain:...) reference in the config file. Keys never appear in shell history or LLM conversation transcripts.
• Allowlists and blocklists: A commands.allowed or commands.blocked list in the config restricts which commands can run, globally or per context. It’s useful when you want an agent to have read access but not delete access.
• Validation before sending: Every command validates its inputs against a JSON Schema before making any network request. Pass --dry-run to validate without sending anything.
• Destructive commands require confirmation: In a terminal, they require you to approve. In a non-interactive session (where agents run), they don’t execute unless you pass --yes explicitly.
• Output can be trimmed: Use --output-fields to return only the fields you need, or --output-template for a Mustache-style template. Smaller responses mean smaller context windows for agents.

The CLI also ships a few helpers for Elasticsearch patterns that have a learning curve:

• elastic es helpers bulk-ingest: Load from a file, directory, or stdin (NDJSON, JSON array, or CSV), with batching, concurrency, and retries handled for you.
• elastic es helpers scroll-search: Stream a large result set as NDJSON with paging handled automatically.
• elastic es helpers watch: Print new documents to stdout as they land — useful for piping into logging tools.

Elastic Agent Skills now use the CLI as their transport

Elastic Agent Skills teach AI coding agents (e.g., Claude Code, Cursor, Copilot, and Gemini CLI) how to work with Elasticsearch and Kibana correctly, including which fields to look at, how to stage a reindex, and which cluster health signals matter. They previously had to hardcode transport details like curl commands with embedded API keys and hostnames. Those break when credentials rotate or the endpoint changes.

Skills now use the CLI. Operations are written in HTTP shorthand in the skill body, and a table at the end maps each to a CLI command:

GET /{index}/_mapping -> elastic es indices get-mapping --index '<index>'

POST /_query -> elastic es esql query --format tsv --query "<esql>"

The skill provides the knowledge. The CLI provides a validated, credential-safe, and policy-scoped way to act on it. Install both with two commands:

npm install -g @elastic/cli

npx skills add elastic/agent-skills

Skills for Elastic Cloud onboarding, Elastic Workflows, and Kubernetes investigation are available now. ES|QL, ingest, reindex, index design, Kibana dashboards, and alerting skills are in progress. The full list of safety rails and the serverless command profile are in the blog post.

Precompute the facts once, so your agents stop rereading the same documents

Standard RAG retrieves full documents into the agent context every time a question comes in. The same source content gets loaded repeatedly, and the token cost adds up with every query. This post from Kathleen DeRusso and team walks through an alternative using Elasticsearch AI Indices.

The idea is to run a one-time batch pass over your document corpus using a Kibana Workflow. For each document, an LLM extracts a Knowledge Indicator (KI): a compact, grounded summary containing a title, a plain-language summary, a list of entities, the questions the document can answer, and a tagline. These get written to an AI Index.

When an agent gets a question, it runs a hybrid BM25 and semantic search over the KI index using ES|QL and grounds its answer in the retrieved indicators. There are no full documents in context.

The numbers from one test query ("What was the actress who played Torvi from Vikings also known for?"):

| Metric | Standard RAG | With KIs | | Tool calls | 8 | 2 | | Tokens used | 386,187 | 27,625 | | Latency | 44.86 s | 15.22 s |

The answers were equivalent and grounded in the same source material. The approach works with LangChain, Elastic Agent Builder, and Claude Code without changes to the agent harness.

The Workflow YAML, the full KI structure, and the complete ES|QL query are all in the post linked above.

See what your LLM calls are doing right in the APM trace view

If you instrument a generative AI (GenAI) application with OpenTelemetry, LLM calls now show up in the Elastic APM trace waterfall as proper spans. A new GenAI tab in the span flyout lets you read the system prompt, the input messages, and the model response for each call with copy buttons for each field. Token counts appear as badges on each GenAI span row in the waterfall. In an agentic trace with multiple LLM calls, you can see at a glance which call is driving token consumption without opening each span individually.

The feature follows OpenTelemetry GenAI semantic conventions (v1.37.0+) and works with any OTel-instrumented provider: OpenAI, Anthropic, AWS Bedrock, and others. No Kibana configuration is needed if your instrumentation already emits GenAI attributes. It is in technical preview on Serverless now, with Stack 9.6 support coming next. Setup instructions and a walkthrough of the OTel attributes are in the post.

13 million tool calls: How we audit AI coding agents across 1,100 machines

AI coding agents like Cursor run under developer accounts. When they execute a shell command, read a file, or call an MCP server, it looks identical to the developer doing the same thing in your endpoint detection and response (EDR) logs. There is no way to tell after the fact whether a command came from a human or a model.

The Elastic Security Labs team built a lightweight auditing setup using Cursor's built-in hooks. Hooks fire at defined lifecycle points (e.g., before/after shell execution, before/after MCP tool calls, file reads and edits, and session start and end). A 280-line dependency-free bash script captures each event as structured JSON and Elastic Agent ships it to Elasticsearch via a custom logs integration.

After rolling out to 1,100+ machines, they had 13 million events. A few things the data showed include:

• File reads outnumber shell commands roughly 4:1. Agents spend most of their time reading the codebase before doing anything.
• More than 300 distinct MCP servers appeared in the data, and 86% of those connected to only one or two people, which makes inventory and risk assessment much easier.
• ES|QL makes it straightforward to hunt for specific patterns: credential file reads, download-and-execute chains, or a full MCP server inventory across the fleet.

The hook approach is sensor-only by default (agents are not blocked, just observed), but Cursor supports blocking hooks if you want to add approval gates for sensitive actions. The same mechanism applies to Claude Code and other agents that expose similar hook interfaces.

The collector script, hooks configuration, and Elastic Agent integration setup are in the repo linked from the post, along with the ES|QL queries, field-level security setup, and notes on privacy and transparency.

Blogs, videos, and interesting links #

• Vector Database:  Dustin Coates introducesElasticsearch Vector Database , a new serverless offering optimized for vector based applications.
• Agentic SOC: Find out how we tookAI alert triage from 60% to 92% accuracy with Maggie Musquez.
• Lucene:  Parker Timmins and Martijn Van Groningen walk us throughLucene query rewrite rules that make two string scan queries in columnar mode 2.3x and 1.6x faster.
• Agent Builder:  Jeffrey Rengifo shares three ES|QL tools and two dozen lines of query text that turnAI root cause analysis into something you can check.
• Kibana dashboards:  Teresa Alvarez Soler and Rudolf Meijering demonstrate how to find outwhich Kibana dashboards get used or catch a broken dashboard.
• Time series data:  Mary Gouseti explains how towrite documents with past timestamps straight into Elasticsearch time series data streams.

Check out these videos:

- [Inside Elasticsearch's metrics engine: A technical deep dive](https://youtu.be/3O6RpGKQZcc?si=1gN-nPJpBcvi4Xcr) by JP Hwang
- [SOC case management in Elastic Security 9.5: Templates and queryable case analytics](https://youtu.be/1Ior7O9UKMA?si=-vXuM5rtMFOdYFtg) by Melissa Burpo
- [Elastic investigates your Kubernetes incidents before you do](https://youtu.be/y8LZMd79N8M?si=SqCqWUk_6Odvp28u) by Jen Luther Thomas

Featured blogs from the community:

• Rebuilding Notion’s lexical search reindexer by Aravind Selvan, Calder Lund, and Gaurang Sadekar
• Processing Heterogeneous IoT Logs with Fluent Bit and Elasticsearch by Christophe Chaussat
• I Gave My AI Agent Access to Elasticsearch. Then I Locked the Door Behind It by Baburaj R

Upcoming events #

Elastic{ON} Tour, our one-day conference series bringing the Elastic community together around the world, is back!

New York-area community: Join us on October 8 for a full day of technical sessions, real-world insights, and opportunities to connect with fellow Elastic users and experts. Save your spot for Elastic{ON} New York.

You can also register to join us in:

• Mumbai — September 30
• Amsterdam — October 20
• San Francisco — November 4
• London — February 25
• Singapore — March 23

Want to speak at Elastic{ON}? We’re still accepting talk proposals for select upcoming Elastic{ON} events! We’re looking for 30-minute technical talks covering Elasticsearch, Kibana, search, observability, security, and real-world use cases. Whether you’re new to Elastic or an experienced user, we’d love to hear your ideas.

Join your local Elastic User Group chapter for the latest news on upcoming events! You can also find us on meetup.com. If you’re interested in presenting at a meetup, send an email to meetups@elastic.co.

The release and timing of any features or functionality described in this post remain at Elastic's sole discretion. Any features or functionality not currently available may not be delivered on time or at all. In this blog post, we may have used or referred to third party generative AI tools, which are owned and operated by their respective owners. Elastic does not have any control over the third party tools and we have no responsibility or liability for their content, operation or use, nor for any loss or damage that may arise from your use of such tools. Please exercise caution when using AI tools with personal, sensitive or confidential information. Any data you submit may be used for AI training or other purposes. There is no guarantee that information you provide will be kept secure or confidential. You should familiarize yourself with the privacy practices and terms of use of any generative AI tools prior to use. Elastic, Elasticsearch, and associated marks are trademarks, logos or registered trademarks of elasticsearch B.V. in the United States and other countries. All other company and product names are trademarks, logos or registered trademarks of their respective owners.