kglite-visual: see a knowledge graph, and let an agent drive it

kglite-visual is an interactive viewer, a headless renderer and an agent interface for .kgl knowledge-graph files produced by KGLite. One command opens a browser on a localhost server; the same binary draws an image without one; and while the server runs it speaks the Model Context Protocol, so an agent can drive the window you are looking at.

The Python wheel has no required runtime dependencies: the graph engine, the HTTP server and the WebGL frontend bundle are all inside one compiled extension. No Node, no separate server process, no database service.

Quick Start

The released package:

pip install kglite-visual==0.1.9
kglite-visual graph.kgl

Open the downloadable 17-node team sample from Getting started. The first exploration takes it through browse, inspect, query, filter, calculate, save and export.

What you land on is a type-level meta-graph, rather than every instance in the source. .kgl files reach 100M+ nodes, so you inspect counts first and load a bounded slice only when it is useful.

Two more commands, no server and no browser in either:

kglite-visual render graph.kgl --meta -o schema.svg      # one image, one JSON line
kglite-visual export graph.kgl --format gexf -o out.gexf # one file for somebody else's tool

Then hand the running server to an agent. The MCP endpoint is on the same port, and its URL is printed on the same stdout line as everything else:

{"url":"http://127.0.0.1:54137/","port":54137,"pid":69850,"graph":"graph.kgl","mcp":"http://127.0.0.1:54137/mcp"}

→ Getting started · Agents and MCP.

What makes it different

Three things this does that a graph viewer normally does not.

An agent drives the window you are watching. The running server speaks MCP at /mcp — no second process to start, no discovery file, nothing to install: attaching an agent is pointing it at a URL. Twenty-seven tools act on one shared view, with ordered changes and optional revision checks: read what is on screen, put a Cypher result into it, expand or collapse, highlight, zoom, recolour, re-lay-out, export it, draw a picture of it, and run the queries you saved under the names you chose. Whoever changes the view — you, the agent, a curl — every connected window sees the change immediately. Watching an agent expand a type and zoom to what it found is the feature, not a side effect of it. → Agents and MCP.

The honesty model: truncation is drawn into the picture. A graph viewer is a machine for showing you less than there is, and the design position here is that the subset must name itself. The response bound lives in core, not in the UI — a guarantee the client implements is not a guarantee — so a curl, an agent and a second tab all hit the same ceiling. Every bounded answer carries {returned, total, truncated}; a slice carries two of those, because nodes and links share one byte budget and a complete node list can sit beside an incomplete link list. And because an image travels without its response, the banner is drawn into the image, beside three more counts for the other ways a picture can be less than its input: types_shown (a canvas too small for the schema draws the largest types and says top 24 of 98), names_shown (labels that lost their cell keep their circle and lose their name), and folded (a fan too big to read is one wedge saying how big it is). → The honesty model.

Layouts chosen from the graph's own structure — including a real map. A force layout is the right tool for a graph with no discoverable shape and the wrong one for a star, a bipartite result or a schema with disconnected families, which is most of what a real graph hands it. So a neighbourhood is drawn as hop rings, a community-structured graph as packed islands with a quiet boundary round each, and unattached nodes as one labelled grid. And where the nodes carry coordinates there is --layout geo: an equirectangular projection whose longitudes are corrected by the cosine of the data's mid-latitude — so a shelf at 68°N comes out its own shape rather than 2.7× too wide — over the world's real coastline at three scales chosen by how much of the world the frame covers, so a North Sea crop gets the fjords and a world map does not carry 400,000 points nothing can resolve. Nodes with no coordinate go in a labelled tray with a count, never dropped. → Layouts.

From Python, and from a notebook

import kglite_visual as kv

view = kv.show("graph.kgl")     # the same server, in-process
view.url                        # 'http://127.0.0.1:54137/'
view.launch_info                # {'url', 'port', 'pid', 'graph', 'mcp'}
view.close()                    # stops it, frees the port

# Or hand over an in-memory kglite graph — through to_bytes(), never the disk.
import kglite
view = kv.show(kglite.load("graph.kgl"))

In a notebook the returned object renders itself in the cell: a proxy-prefixed iframe where jupyter-server-proxy can reach the port, and no iframe at all — the URL plus an ssh -N -L hint — where the kernel looks remote, because a localhost iframe from a remote kernel is a silently blank frame.

show(path) is the large-graph answer: handing over an in-memory graph costs about 2× the graph's size at the moment of the call. → Python API.

Render and export

The browser’s Export dialog previews visible instances and exact retained relationships, the wider loaded-node induced graph, or a deterministic SVG/PNG. Downloads refuse stale previews after shared changes. The image uses a server layout, separate from the browser camera. Data tables also export scoped CSV.

# an image: --meta, --cypher "…" or --expand type=T rel=R dir=out
kglite-visual render graph.kgl --meta -o schema.svg
kglite-visual render graph.kgl --cypher "MATCH (f:Field) RETURN f" --layout geo --format png

# a file: graphml | gexf | csv | csv-edges | json
kglite-visual export graph.kgl --format gexf -o graph.gexf
kglite-visual export graph.kgl --format csv --cypher "MATCH (n:Field) RETURN n"

The render's layout is seeded and deterministic — the same request produces the same bytes, forever — so --seed is how you get a different arrangement of the same data, and an exact golden baseline is possible at all. Each command writes its file and prints one JSON line describing it; nothing else ever touches stdout. → Render · Export.

Requirements

CPython 3.10+ (one abi3 wheel serves every version from 3.10 up) on macOS, Linux and Windows, plus a .kgl file supported by the embedded KGLite reader — this version pins kglite 0.21.2, which reads compatible files written by older releases. Building from source additionally needs a Rust toolchain; the published wheels and the source distribution both carry a prebuilt frontend, so neither needs Node at install time.

Documentation

Full docs at kglite-visual.readthedocs.io.

• Getting started — install the package and open the sample.
• Your first exploration — browse, inspect, query, filter, calculate, save and export one graph.
• The viewer — the app in full: the honesty model · layouts · query surfaces (editor, saved queries, generated tables, the path builder, PROFILE/EXPLAIN) · appearance and filtering.
• Agents and MCP — the launch contract, the JSON twin, the twenty-seven tools, window.__kglv, and what an agent may claim about a screen it cannot see.
• Python API · CLI reference · Render · Export.
• Concepts — bounds in core · the protocol · storage · memory.

Rendering is cosmos.gl (MIT, OpenJS Foundation): a WebGL GPU force layout, fed over a binary protocol — typed-array buffers for topology and positions, JSON for metadata.

Stability

Alpha, pre-1.0. The launch contract (one JSON line: url, port, pid, graph, mcp), the render and export summary lines, and the MCP tool names are the surfaces to depend on; everything else may move. CHANGELOG.md records what a user can see change.

License

MIT — see LICENSE.