Context graphs for agents
that need to reason.

Q: How is this different from vector search or RAG?

Vector search finds passages that look similar to a question. A graph stores explicit, typed relationships, so an agent can follow on_top_of , acquired , or part_of edges across several steps and keep the evidence attached at every hop. They work well together: retrieve broadly with embeddings, then use the graph to resolve relationships and cite sources. Read how context graphs work .

Q: What kinds of data can it handle today?

Images (photos, screenshots, video frames) with OpenGraph Image Documents ( .txt , .md , .pdf ) with OpenGraph Text Tables (CSV, TSV, Excel, and photos of tables) with OpenGraph Table, in preview Audio and video files aren't read directly yet. For video, sample frames into a folder and build an image graph from them.

Q: Can I contribute or request a feature?

Yes. Issues and pull requests are welcome on GitHub . New test images, extraction edge cases, and documentation fixes all help. To talk an idea through first, join the Discord or email team@opengraphai.io .

Turns multimodal data into queryable knowledge graphs. Agents can traverse relationships, reason across sources, and cite the evidence.

Try the playground Read documentation

OpenGraph AI works with coding agents.

Build graphs, inspect what each file contains, and ask questions answered from the graph with cited evidence.

Give your coding agent a connected view of your workspace instead of another pile of isolated search results.

More context.
Less overhead.

Works inside your agent.

Prompt naturally. OpenGraph builds and queries the connected context behind the answer.

Example output

head_rgbep03_step-025.jpg Where was the mouse?`mouse`*next_to*`tablet`

Robotics

Give robots a scene memory they can query.

Record what was on, inside, or next to what.
Ask where an object was without replaying video.
Trace every answer to the camera frame behind it.

Explore robotics

person holding umbrella ./media

personholdingumbrella4 matches

street-03.jpg0.96
street-11.jpg0.94
street-14.jpg0.91
0:41

t00042.jpg0.88

Image & Video Search

Find the moment by how things relate.

Match on relations, scenes, attributes, and visible text.
Search “a person holding an umbrella.”
Sample video into frames and search those too.

Explore image & video search

Which sites were marked damaged, what does the manual say to do, and do the photos show it?

opengraph-table query_graph north_yard status “damaged”site-reports.xlsx, Sites, row 7
opengraph-text query_graph “Replace cracked roof panels within 48 hours of inspection.”maintenance-manual.pdf
opengraph-image get_image_entities cracked_panel 0.88north-yard-04.jpg

North Yard is the only site marked damaged¹. The manual says to replace cracked panels within 48 hours², and north-yard-04.jpg shows one³.

Multimodal AI Agents

One agent that reasons across every format you have.

Run the image, text, and table servers side by side.
Let the agent pick the right graph for each step.
Cite the file behind every part of the answer.

Explore multimodal agents

4.2 Results

Electrolyte additives and cycle life in LFP cells

Cells were cycled at 1C between 2.5 and 3.65 V at 25 °C, twelve cells per batch. Both batches started at full capacity. After 500 cycles, capacity retention rose from 78% to 91% with the additive, and the gap opened after cycle 200 (Fig. 3, Table S2).

Impedance growth followed the same split: baseline cells added 41% to their DC resistance by cycle 500, against 17% with the additive.

cycle-life-study.pdf

fig-3.pnghas_text “Capacity retention (%)”

Cycle	Baseline	Additive
250	88.2	95.6
500	78.4	91.1

supplement.xlsxsheet S2, row 14

Research

Connect papers, figures, and data.

Link claims, figures, and data tables in one graph.
Explore connections across findings and concepts.
Trace every answer back to the file it came from.

Explore research

Answers that follow the edges.

The agent hops across typed relations to the frame that shows it, instead of guessing from the closest caption.

Example output

Where did I leave my phone, and what was next to it?

Robot camera frame ep03_step-010: a phone and a glass of tea on a kitchen island beside a laptop, under a wall clock that reads 09:41. — ep03_step-010.jpg

phone on_top_of kitchen_island
phone next_to glass_mug
phone seen_in ep03_step-010.jpg

Every link carries a score.

Extraction records how sure it was about each node and edge, from 0 to 1. Raise the bar to see which links hold up.

Example output

Robot camera frame ep03_step-040: a laptop, potted plant, mouse, and fruit basket on a kitchen counter, with a tablet, clock, kettle, and coffee machine behind them. — ep03_step-040.jpg1600 × 1067

Minimum confidence 0.60

7 of 8 objects and 6 of 8 relations kept. Weakest relation kept: digital_clock behind tablet, 0.64.

Every node points back to its source.

Pick a source type to see the box, sentence, or cell a node was extracted from.

Example output

Robot camera frame ep03_step-025, with a box drawn around the fruit basket on the counter. — ep03_step-025.jpg1600 × 1067

type: "object"
label: "fruit_basket"
bounding_box: [0.71, 0.49, 0.29, 0.31]
confidence: 0.95
seen_in: ["ep03_step-010", "ep03_step-025", "ep03_step-040"]

A4fxCorvane Robotics

	A	B	C
1	Supplier	Region	Renewal
2	Halden Components	EU	2026-03
3	Pellworth Supply	US	2025-11
4	Corvane Robotics	US	2026-07
5	Kessel Drive Systems	EU	2026-01

ContractsRenewals

type: "entity"
name: "Corvane Robotics"
entity_type: "organization"
source_cells: [{"row_index": 3, "col_index": 0}]
confidence: 0.93

From raw context to a queryable graph.

Start from the terminal, connect your sources, and query the same graph from the interface your agent already uses.

CLI

Build, inspect, and query a graph from your terminal. Good for scripts, CI, and a first look at your data.

Questions, answered.

What OpenGraph AI does today, where your data goes, and how to start building.

Still have a question? Ask us directly or dig into the docs.

Talk to us Read the docs

What is OpenGraph AI?

OpenGraph AI is an open-source context layer for AI agents. It turns your files into knowledge graphs of entities, the relationships between them, and the source each one came from, then gives agents tools to query those graphs over MCP or a CLI.

It ships as separate packages per data type. OpenGraph Image and OpenGraph Text are on PyPI; OpenGraph Table is in preview.

How is this different from vector search or RAG?

Vector search finds passages that look similar to a question. A graph stores explicit, typed relationships, so an agent can follow on_top_of, acquired, or part_of edges across several steps and keep the evidence attached at every hop.

They work well together: retrieve broadly with embeddings, then use the graph to resolve relationships and cite sources. Read how context graphs work.

What kinds of data can it handle today?

Images (photos, screenshots, video frames) with OpenGraph Image
Documents (.txt, .md, .pdf) with OpenGraph Text
Tables (CSV, TSV, Excel, and photos of tables) with OpenGraph Table, in preview

Audio and video files aren't read directly yet. For video, sample frames into a folder and build an image graph from them.

Which agents and tools does it work with?

Any MCP-compatible client, including Claude Code, Claude Desktop, and Cursor. Register a server once and the agent picks the right tools on its own:

claude mcp add opengraph-image -- opengraph-image-mcp

There is also a CLI for scripts and CI, and every build writes a plain graph.json you can load in your own code.

Does my data leave my machine?

Building and querying a graph call your model provider, currently Claude through the Anthropic API, so source content is sent there for extraction. The graph itself is written to a local opengraph-out/ folder next to your data, and you decide where it goes from there.

Check your provider's data terms before processing anything sensitive.

What does it cost? Do I need an API key?

The packages are free and open source under the Apache-2.0 license. Extraction and querying use Claude, so you need an Anthropic API key set as ANTHROPIC_API_KEY, and that usage is billed by Anthropic. Build cost grows with the number of files you process.

How do I check where an answer came from?

Every node keeps a link to its source. Image objects list the images they were seen in, text claims link to the document that states them, and table values keep their file, sheet, and cell.

Ask the agent to show its evidence, or call get_image_entities or get_document_entities to inspect everything extracted from a single file.

How do I get started?

Install the image package and build a graph from any folder of pictures:

pip install opengraph-image
opengraph-image build ./photos

The quickstart walks through the rest. To see a finished graph first, open the live robot-memory sample; it needs no install.

Can I contribute or request a feature?

Yes. Issues and pull requests are welcome on GitHub. New test images, extraction edge cases, and documentation fixes all help. To talk an idea through first, join the Discord or email team@opengraphai.io.

The open context layer for reasoning agents.

OpenGraph AI turns tables, text, images, audio, and video into queryable knowledge graphs, so agents can reason over the relationships in your data and retrieve the evidence behind an answer. The first packages cover images, documents, and tables.

Open by design.

One context layer, designed to fit the stack you already trust.

Yours to run. Apache-2.0 packages that run on your machine. The graph is a local JSON file you own.
Connect every signal. Images, documents, and tables each build a graph today, with every node tied to the file it came from. Audio and video are next.
Fits your stack. graph.json loads straight into NetworkX, so it drops into the pipelines and notebooks you already run.
One interface for context. Agents reach the graph through MCP tools, the same way they reach everything else.

Talk to us

Building something on top of OpenGraph AI? Tell us about it. We read every message and reply.

Context graphs for agents that need to reason.

Works inside your agent.

Robotics

Answers that follow the edges.

Every link carries a score.

Every node points back to its source.

Open by design.

Talk to us

Context graphs for agents
that need to reason.