The Sovereign

A research lab that never stops

Open source and self-hosted. Give it a topic: every night it reads the web and the literature, keeps each fact with its source, and proves what it learned on an exam.

Start your lab $git clone https://github.com/cpldxx/the-sovereign && cd the-sovereign && ./sovereign up

An example learning cycle, playing out

Domain: Solid-state batteries Example

How your lab learns

One learning cycle every night, at 03:00 by default, or whenever you start one.

1

It asks what you want to know

The Head, the agent that runs your lab, interviews you and writes a research charter: the goal, the core questions, what to leave out. Every run steers by it.

One charter per topic, yours to edit in Settings.

Research charterExample
Goal
Know what stands between solid-state cells and mass production.
Core questions
Which electrolytes lead, and why? What limits yield? Who runs pilot lines?
Leave out
Consumer product rumours.
Depth
Engineering detail, cited.
2

It reads the web and the literature

It searches, opens what it finds, and reads papers from OpenAlex, arXiv and PubMed. Pages behind robots.txt, a login or a paywall are left alone.

47pages read in one night: four research missions and a check on the week's news
Real run, 11 Nov 2026
Tonight's readingExample
  • PaperA review from arXivread
  • PaperAn open-access article, PubMedread
  • OfficialA standards body's noticeread
  • NewsTrade pressread
  • BlogAn engineer's notesread
  • SiteKeeps bots outskipped: robots.txt
3

It keeps every fact with its source

Pages become facts in a knowledge graph. More sources make a fact stronger, a newer report replaces an older one, and what never holds up is forgotten.

1,010facts stored that night, each with the page it came from
Real run, 11 Nov 2026
New in the graphExample
  • Sulfide electrolytes reach high conductivity3 sources, strong
  • Lithium-metal anodes grow dendrites2 sources
  • Dry rooms raise cell cost1 source: a weak signal
  • A pilot line targets 2026replaced by a newer report

What you get back

A score you can watch, answers you can check, and a graph your other agents can query.

4

It proves what it knows

An exam written from sources the research is never allowed to read, answered from the graph alone. What it can't answer becomes the next night's research.

35of 100 on a 30-question exam, answered from its own graph
Real run, 11 Nov 2026
Exam, 30 questions7 answered, 7 in part, 16 open
5

It answers with sources, and says how sure it is

Ask the Head anything about the topic, by text or by voice. It searches the graph, reads the facts behind it, and tells a confirmed fact from a weak signal.

25 tools the Head can use: the graph, the web, papers, its team.

HeadExample
YouWho is closest to mass production?

Two makers run pilot lines source 1 source 2; one has moved its target to 2027 source 3.

Confirmed by three sources. The newer date replaced an older report.

1 Official2 News3 Official

Your other agents can use what it knows

The graph is an MCP server. Make an API token, connect Claude Code or any MCP client, and it can query what your lab has learned.

15 read-only tools over MCP, one token per person.

.mcp.jsonfor Claude Code
{ "mcpServers": { "sovereign": {
  "type": "http",
  "url": "http://localhost:8000/api/kg/mcp/readonly",
  "headers": { "Authorization": "Bearer svk_…" }
} } }
query_knowledge_graphget_entityread_synthesisexam_resultsdaily_reportsearch_papersread_webpageand 8 more

It gets measurably better

Every number here comes from the lab's own records. The exams use held-out questions it was never taught.

41to55

Research aimed at its mistakes works

After research missions on the practice questions it got wrong, its score on a separate, held-out exam rose from 41 to 55.

Experiment, test domain
22to35

The same exam, sitting after sitting

On the current 30-question exam: 22 at its first sitting, 35 on 11 November. The questions stayed the same.

Real runs, Nov 2026
9× faster

Tuned for a model you run yourself

With reasoning off, an extraction on a local 35B model takes about 19 seconds instead of 171, at the same quality.

Measured on qwen3.6:35b

Scores move both ways from night to night. A sitting the model couldn't grade is called off, never counted as a zero.

Free on your machine, or on Claude

Pick at the first start, switch any time. Either way the graph stays on your server.

Local models

Free and private: nothing leaves the machine but its searches and page visits.

  • Modelqwen3.6:35b on Ollama
  • Memoryabout 24 GB free (a 64 GB Mac is comfortable)
  • Costnone, no API key
./sovereign up

Claude

Any machine, no GPU. The agents think on Claude; only a small embedding model runs locally.

  • Modelclaude-opus-5-5, or any Claude model
  • Memoryno GPU needed
  • Costper token, with your Anthropic API key
./sovereign up --claude

What else it can do

Each of these runs in the lab today.

  • Scholarly search. OpenAlex, arXiv and PubMed, with open-access full text.
  • Its own search engine. A local SearXNG, or Tavily and Brave with a free key.
  • Pages that need a browser. JavaScript-rendered sites, and PDFs up to 60 pages.
  • Graph search by meaning. The closest entities and facts, the strongest facts around them, and their sources.
  • Synthesis pages. Per core question: the consensus, the debates, the open questions, the key sources.
  • A briefing every cycle. What it learned, what changed, what needs your attention. It can read it aloud.
  • Voice. Talk to the Head and hear it answer; speech recognition runs on your machine.
  • Many topics, many people. A database per topic; share one as viewer, editor or owner.
  • A nightly security check. No answer without sign-in, no reach across topics, no page past robots.txt.
  • It recovers on its own. After a crash it picks up the pages it had read and waits while the machine is short of memory.

It doesn't act in the world: no messages, orders or bookings. It finds out what is true and explains it.

Start your lab tonight

All it needs is Docker. With local models there is no API key to get.

git clone https://github.com/cpldxx/the-sovereign.git
cd the-sovereign
./sovereign up

Then open localhost:8000, create your account, and name a topic. Models guide, troubleshooting.