Agenten stärken
Installation vom Herausgeber. Kein PlugMCP-Endpunkt.
Host a copy with our keys and logs.
Noch kein paketiertes Release in der Registry. Folgen Sie der README des Repositorys.
Remote-URL oder ein Befehl auf Ihrem Rechner? Worin sich die beiden Installationen unterscheiden
Auf PlugMCP ausführen
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
StartenVom Herausgeber
prax is a self-hosted library for the papers, web pages, manuals and notes one person reads. Its search matches words and meaning at once, its graph records what the documents say about each other, and an AI agent can search and read it through the same interface its owner uses. Everything lives in one SQLite file and a folder of originals, on your own machines, with models you choose.
A self-hosted library for the papers and web pages one person reads: hybrid search, a graph whose links cite their sentence, and an MCP interface for agents
pip install -e ".[serve,work]"
export PRAX_DATA_DIR=~/prax-data
prax serve # http://127.0.0.1:8000/ui/
prax add paper.pdf
prax add https://example.org/article
prax search feedback delay networks
prax ask --answer why do FDNs colour the tail
An empty store works from the first command. A new document is
archived as it arrives, and prax work --watch, run on the machine
with the models, does the rest: its text, title, graph links and
vectors. To keep the parts running, list them under run: in
prax.yaml and run prax up --install, which starts them at login on
Windows, Linux or macOS. Setting up the Zotero import and the browser
extension is in docs/howto.md.
My reading was spread over a Zotero library, an external disk I called
zoetrope, a NAS and a year of browser tabs. I could only find a paper
again if I remembered where I had put it. I wanted one place that keeps
the originals and reads them closely enough to answer "which of my
papers says this, and where". An agent working with me should be able
to search it without being handed whole PDFs. I built prax with Claude
Code over a few weeks in 2026. I decided what it should do and measured
whether it did, and the log of that work is in
docs/log.md. It is shaped by one library and one set
of machines, and it is published so the design and the measurements
can be read and reused.
Say I want to know why a feedback delay network colours the tail of a reverb. A search from the shell returns documents, each with the passage that matched (the brackets mark the matched words) and where in the document it sits:
$ prax search feedback delay networks colour the tail
…
3. The Role of Modal Excitation in Colorless Reverberation doc 9577
[Feedback] [delay] [networks] (FDNs) are a computationally efficient
structure for artificial reverberation…
text · page 1 · THE ROLE OF MODAL EXCITATION IN COLORLESS R…
4. Allpass Feedback Delay Networks doc 13371
…arbitrary connection of [delay] lines, namely [feedback] [delay] [networks]
(FDNs). We present…
prax ask --answer goes further when there is a model on the host. It
searches again with words of its own, reads on in the documents that
looked promising and drops what does not help. Then it writes an answer
whose citations point at the passages it kept. The trail of those steps
is shown under the answer, so you can see why a source was used.
An agent asks the same library through the MCP server, and what comes
back is sized for its context window. The same question asked by an
agent, with three hits, is 1.8 KB of JSON. It opens with the region of
the library the hits live in, and each hit carries a cite link that
still finds the passage after the document is re-chunked:
{"doc_id": 270, "title": "Building the Erbe-Verb: Extending the Feedback Delay Network Reverb for Modular Synthesizer Use",
"snippet": "…on a 4-[delay] [feedback] [delay] network [reverb] (FDN) as proposed by…",
"heading": ["…", "2. BASIC DESIGN"], "page": 1, "published": "2015",
"cite": "#doc/270?chunk=1757785&find=a+unitary+feedback+matrix"}
The agent can then ask how two things are connected. Asked for
"feedback delay network" and "Schroeder reverberator", connect
answers in 1.2 KB with one path of two steps through a paper on
scattering delay networks, and each step quotes the sentence it rests
on. One says the paper uses the Schroeder reverberator: "Starting with
the Schroeder reverberator [3, 4], a wide variety of approaches for
room acoustics simulation have been introduced…". The other says it
uses feedback delay networks, on the strength of "…found that they
perform better than Feedback Delay Networks (FDNs)". That is a generous
reading, and because the quote comes with it, an agent or a person can
see so and weigh it.
Most PDF tools stop at the text. prax pulls each figure out, or renders the region above its caption when the figure was drawn with vector paths. A vision model then reads it together with the paragraphs that refer to it. That is why a plot comes back as "the frequency responses of the five learned CNN kernels against the ground truth filter" instead of "six stacked curves". The reading is a passage of the document, so you can search for a figure by what it shows, cite it, and an answer can use it. A scanned book with no text layer is read the same way, page by page.
A model reads each document against a small ontology: which paper uses which method, which manual belongs to which piece of gear, which recipe needs which ingredient. Every link it writes records the document, the sentence, the model, the batch it ran in and the ontology version. I wanted that because models misread papers, and a link I cannot trace to a sentence is one I cannot check. When a later reading disagrees, the old link is given an end date and kept. A database trigger refuses any attempt to delete a link or change what it says, so you can ask what the graph said on a given day. A whole batch from one model can be withdrawn without touching anyone else's work. A fact that states its own date ("she worked at the lab from 2019") keeps that date apart from the date prax learned it. Names are kept apart from things: apple the ingredient and Apple the company are two entities, and a merge of two names into one thing can be taken back.
Keyword (SQLite FTS5) and vector search run over the passages and over a short description of each document. The four lists are merged per document, and you can filter them by kind of document or by subject. Acronyms the library defines ("feedback delay network (FDN)") are expanded on the keyword side. Each hit shows when its document was published, as precisely as the source says, so you can ask for only what appeared after a given year. A document that a newer one replaced is still found, a few places lower, with a note naming its replacement. The numbers, with their caveats, are further down.
Notes, project logs and write-ups are Markdown pages with a revision history, and they are searched and read like any other document. A link from a page to a document becomes a link in the graph. If you put a question between two comment lines, prax answers it there with sources and answers again when new documents arrive. The second time, the model sees its earlier answer and is told what is new, so it revises. If you edit inside an answer, prax leaves that block alone. A daily briefing page lists what arrived and which answers changed.
prax import reads exports from GitHub stars, chat apps, bookmarks, Pocket,
Raindrop and Medium, and prax sync sends a project's notes from its
git working copy.prax heal lists recurring damage (a placeholder entity, a page
captured twice, a scan filed under its cover's title, figures not yet
read) and prints the command that fixes each. A fix hides a wrong
document with its history and deletes nothing.docs/design/BRIEF.md).The web UI at /ui/ has search and ask, a document with its context
and figures, the graph, the review queue, your pages, the inbox and the
jobs. The prax command covers the whole library from a shell, and it
is how I use it most:
prax where things stand, what to type next
prax search granular synthesis find documents
prax ask --answer how does a feedback delay network work
prax add ~/Downloads/paper.pdf a file, a folder, a URL, or piped text
prax import links bookmarks.html what a service exported
prax show 4312 | less read one in the terminal
prax graph "wave digital filter" what the graph knows around a name
prax status · jobs · heal · backup · doctor
prax up · serve · work --watch run it (--tray: an icon in the tray)
A command is one HTTP call to the service, so --json makes any of
them usable from a script and --door points the same command at
another machine. The browser extension sends what you are reading,
including pages behind your login. The MCP server gives an agent the
same reads as the UI, each as one HTTP call. They cover search, a
document by offset and length, the graph around a name, how two things
connect, what changed in a week, and a figure or a scanned page as an
image. Whatever a script or an agent writes carries its name, so you can
inspect it or take it back as a unit. Recipes and the contract are in
docs/integrating.md.
Which model does which step is a line in prax.yaml. It can be a GGUF
served by llama.cpp on your own card, any OpenAI-compatible server, or
the Claude API for the few documents worth paying for, and where a
person does the job better the line says none. A paid model is called
only by a command you run, so private material stays on your machines unless you send it. The
library below was read by local Qwen models on one RTX 4090, and the
comparisons with the hosted model are in docs/eval/.
prax runs on one machine or two. In the two-machine setup a small board holds the database and the service, and a desktop with a GPU does the model work through the same HTTP interface. The serving path is held to a gigabyte of memory so that the board can be a Raspberry Pi or an N100 box.
These are the numbers of my own instance on 7 October 2026. It holds a Zotero import, the contents of a NAS and a year of browser captures, after a few weeks of model passes. They show the scale prax has been run at; they are not a benchmark.
Retrieval is measured on 62 questions I wrote about my own library
(keyword, paraphrase and structure questions), as the mean reciprocal
rank of the document I had in mind. On 8 September 2026 keyword search
alone scored 0.82, vector search alone 0.80 and the merged search
0.79, so merging did not help yet
(retrieval-library-2026-09-08).
After a month of changes the merged search scores 0.90, with the right
document first for 87% of the questions
(retrieval-facts-list-2026-10-05).
Those changes were checked against the same 62 questions, so 0.90 is
an optimistic figure. A held-out set of questions is still to be
written, and the results can be repeated only on this one library.
prax is for one person who wants to keep the originals, see where every claim in the graph came from, and give agents bounded access. If one of these describes you better, another tool is the better choice:
world_at), but under 1% of its facts are dated so far, and its
facts always link two things.docs/compared.md sets out how prax stores facts,
time and evidence beside Utopia, Graphiti, Cognee and Basic Memory, as
read from their code at named commits. The wider survey is in
docs/research.md.
A document's bytes are archived once under their SHA-256, so the same file sent twice is one document, and the database keeps the metadata and the hash. A parser writes the text, which is stored under its own hash and stamped with the parser that made it. The text is cut into passages that each know their place in it: text under a heading, tables, figures, display equations, code, the entries of a reference list. The passages go into an FTS5 index and a usearch vector index. Then a model reads the document against the ontology modules it belongs to. Each relation it finds becomes a link with its evidence, or an item in a review queue when it fits no type.
Everything derived from a document (text, passages, vectors, links, summaries, figure readings) is a model's work and is stamped with what produced it. When a better model or prompt comes along, the documents read under the old stamp can be found and read again, and the earlier reading stays beside the new one. The originals are the only thing that cannot be made again, which is why they are the only thing kept as they arrived.
The service is the only process that writes to the database. Parsing,
extraction and embedding are batch jobs, done by a worker that fetches
work from the service over HTTP and posts the results back. That lets
the worker run on the GPU machine while the database stays on the
board.
docs/architecture.md follows a document and
a query through the code.
prax is one person's tool. There is no installer and no multi-user
story, and the defaults reflect one library. The service has not yet
moved onto the serving board, though the code for it is in
deploy/, and what else is unfinished is in
docs/PLAN.md. Every push runs the test suite on Linux
and Windows and the quick start above in a fresh venv
(ci.yml,
scripts/smoke.sh). The extension has a test bed
that drives it in headless Chrome and Firefox
(scripts/extension_bed.mjs), and I use
it daily in Firefox, Waterfox and Chrome. Issues and pull requests are
welcome, though they may wait.
The service is meant for a private network (a LAN or a VPN such as Tailscale). Its administrator token can read and change everything, the named tokens you hand out see only what you give them, and nobody has audited it. Do not expose it to the internet.
The design notes behind particular parts are
stratification,
identity,
generalizing,
normalization,
review,
packs,
communities,
graph-files,
meta,
ui,
claude-workflow and the
design brief. How the ontology grew is in
v2, v4,
v5, v6,
v7, v8,
v9, studio,
electronics and
craft.
The praxinoscope came after the zoetrope: the same spinning drum, with mirrors that gave a sharper image. prax came after zoetrope, the external disk that held the same library before.
MIT, see LICENSE, except the browser extension.
clients/browser-extension/ is AGPL-3.0
because it bundles SingleFile for page snapshots, the way the Zotero
connector does; it is a separate program that talks to the service over
HTTP. The test fixture under tests/fixtures/ holds
open-access papers under their own Creative Commons terms, listed with
their licenses in
tests/fixtures/zotero/README.md.
A self-hosted library for the papers and web pages one person reads: hybrid search, a graph whose links cite their sentence, and an MCP interface for agents
Tools
Noch keine Tool-Liste. Remote-Server erhalten eine nach dem ersten Health-Check.
FAQ
Kopieren Sie auf dieser Seite die stdio-Konfiguration für Cursor, Claude Desktop, Claude Code oder VS Code. Sie nutzt die Installation des Herausgebers, keinen PlugMCP-Proxy.
In diesem Eintrag ist keine erforderliche Umgebungsvariable aufgeführt.
Prax läuft lokal über stdio.
Zuletzt geprüft: 2026-11-14. A self-hosted library for the papers and web pages one person reads: hybrid search, a graph whose links cite their sentence, and an MCP interface for agents
