Nine small instruments on one bench: ask, browse, remember, schedule, consult, drive the desktop, capture the screen, read the materials, read the other pages.
Nine serversstdio JSON-RPCPython 3.10+Runs beside Quantav1.0.0
What an MCP server is
A small program on the other end of a pipe.
Each server here is a short Python program speaking stdio JSON-RPC: one JSON object per line, stdout kept JSON-only so nothing but protocol ever goes out. Quanta starts it, reads its tool list, and calls a tool like any other step in a turn.
They import Quanta's config and read the app data directory, so they run beside a Quanta checkout rather than as standalone services. That is both a constraint and the point: the server sees the same files, memory and pages the app does.
One line, one message
No framing, no length prefix, no HTTP. A newline is the boundary; a stray print breaks the pipe.
Small on purpose
Each server does one thing and exposes a few tools. When one misbehaves, you can read the whole file in a minute.
Reads the app's data
Memory notes, scheduled runs, page outputs and materials are read from the same directory the window uses.
SkillA written procedure plus a script. Loaded when its description matches the request.SKILL.md
PluginCode inside the agent runtime. Changes how a run behaves rather than what it can call.dsh ACP
ConnectorA door into the agent from somewhere else — a chat app, a microphone.bridge
MCP ServerA separate process exposing tools over stdio JSON-RPC.stdio
The bench
Nine instruments, one bench.
Pick a server on the left. Each scene is a front-end simulation of what it does inside a turn — the buttons work, the API does not.
mcp · v1.0.0
stdio json-rpc one json per line
quanta-ask
Asks the user a question through Quanta's selection-card UI, and waits for the answer instead of guessing.
Three steps, one owner each. Step two blocked on the API key; the crawler page already has a token that can be reused.
Failure clustering — Sep
analyse · yesterday 21:40 · 4 clusters
Timeouts cluster around two hosts. The third cluster is a cookie wall, not a network fault.
Run log · 128 new
crawler · yesterday 23:58 · partial
Two platforms complete, Bilibili stopped at page 3 with 412. Cookies are still valid until Friday.
Negative share, week 41
sentiment · today 07:00 · 11%
Down from 18%. Most of the drop is one thread cooling off, not a change in tone.
One by one
Nine entries, each with its own argument.
The bench above is for poking. Below, each server is on its own: what it does, what it will not do, and the zip.
MCP-01quanta-ask
When the answer is not in the model.
Some decisions are not the agent’s to make: which folder, which account, which of two nearly identical files. quanta-ask turns that moment into a card in the window — a few options, one choice — and holds the turn until a person answers.
The turn waitsNo default, no guess, no answering its own question. The tool returns when the card does.
Selection, not free textOptions come from the caller, so the answer is always something the code can act on.
It stays in the transcriptQuestion and answer remain in the thread, where the next turn can read them.
assumed — saved to ~/Downloads/quanta-export, because that folder was used last time.
wrong foldernobody was asked
ASKWhere should the export go?
~/Desktop/quanta-export
~/Documents/exports/2026-10
the working folder
answer · ~/Desktop/quanta-export
Two ways to end the same turn. The first guesses and is sometimes wrong; the second waits and is occasionally slow.
MCP-02quanta-browser
A tab the agent can actually see.
Quanta’s browser is a real tab inside the window, not a headless browser in a container. quanta-browser drives that tab — open a URL, click a link, read the page — and what it read stays on screen where you can check it.
The tab you can seeNo second, invisible browser with its own cookies and its own version of the page.
Read, then quoteThe text comes back as content, so an answer can point at what is on screen.
Closing ends itNo background automation is left running after the turn.
Three tabs, three kinds of page. What differs is what comes back: text, links, and how much of it was worth quoting.
MCP-03quanta-memory
Remembering is a file operation.
The memory store is a folder of Markdown notes with frontmatter, filed by area. quanta-memory is how a turn touches it: save a note, search by meaning, read one back, link two together, list what is there.
Plain filesEvery note opens in any editor, and the folder can go into git.
Filed by areaNotes sit under the page they came from, not in one undifferentiated pile.
Tools, not stuffingsave · search · read · link · list are calls the agent makes, not context pasted in.
When the answer has more than three fields, use a table; keep the headers Chinese.
wrote · memory/Wiki/summaries/m-20261020-7f2a.md
Prefers tables in Chinese
Wiki/summaries · m-20261002-a1b2c3 · chat
When the answer has more than three fields, use a table.
Crawler needs a login for Bilibili
Raw/sources · m-20260928-k2 · crawler
Without cookies the comment endpoint returns 412 after the first page.
2 of 84 · on-device index · ranked by meaning
Crawler needs a login for Bilibili
Raw/sources · m-20260928-k2 · crawler
Without cookies the comment endpoint returns 412 after the first page.
links · m-20260914-c7, m-20261002-a1b2c3
Save, search, read — three of the six tools. Nothing here is a black box: every result is a file you could have opened yourself.
MCP-04quanta-schedule
Work that happens while the window is closed.
Scheduled runs are how Quanta keeps working when nobody is watching. quanta-schedule lists them, switches them on and off, and can start one immediately — the same entries that appear on the Scheduled page.
Cron, not alarmsRecurring work is a schedule with an owner, not a one-off reminder.
The same sandboxA run gets the tools and the folder the session already had; being unseen widens nothing.
A run leaves a recordOutput lands on the page it belongs to, so the next turn can read it.
The schedule is the boring part, and that is the point: three entries, three rhythms, one place to switch them off.
MCP-05quanta-harness
A second opinion, on demand.
When a decision is expensive, one model’s opinion is not enough. quanta-harness puts the same question to Claude Code, Codex or DeepSeek as consultants — separate contexts, no edits — and brings their answers back to the main thread, which still decides.
Consultants, not delegatesThey answer the question and return. The plan stays where it was.
Separate contextsEach consultant gets the question, not your whole history.
Disagreement is the outputTwo objections are worth more than three agreements, and the transcript keeps both.
Cache the parse, not the render. The parse is the expensive half.
1 call · read-only · 38s
codex
consultant · read-only
Agrees on caching, but the invalidation key is wrong — it ignores locale.
1 call · read-only · 51s
deepseek
consultant · read-only
Two objections is a signal. Measure first; the cache may not be the bottleneck.
1 call · read-only · 44s
The value is not a vote; it is seeing where the answers stop agreeing, while the plan stays in one pair of hands.
MCP-06quanta-cua
Hands on the desktop.
Some work only exists as a desktop app with no API. quanta-cua moves the pointer, clicks, types, and then looks at the screen to see whether it worked — through @trycua/cua-driver, on your machine.
Move, click, lookThe loop is not optional: every action is followed by a look at what it did.
Needs a driver@trycua/cua-driver, plus QUANTA_DATA_DIR and QUANTA_CUA_ROOT.
Your desktop, your riskIt can press anything you can press, which is why the permission prompt stays in the way.
No click yet. The move is its own step, so a hover menu can open first.
click · left · Settings
One press, then a look — before the next press, not after the next three.
screenshot · window matched · 1440×900
The window on screen is the receipt: same title, same three rows.
Three states of one loop, drawn around the same desktop. The screenshot at the end is the only proof the click worked.
MCP-07quanta-screen
When it can only look, not touch.
When the driver is missing — another machine, a locked-down system — quanta-screen is the smaller promise: take a screenshot, crop what matters, and hand the image to the model. Look, but do not touch.
Fallback by designIt exists so a machine without the driver still has eyes.
Crop before sendingA window crop is smaller and more legible than a whole desktop.
No inputA screenshot cannot click. If the task needs a click, it needs cua.
Everything, including the parts that were never going to matter.
window crop · 720×420 · 96 KB
One app, its title bar, and the three rows that actually changed.
crop · 512×288 · 41 KB
Just the field that failed, at a size a model can read without guessing.
Same screen, three crops. Sending less is often the difference between a model that reads the window and one that guesses.
MCP-08quanta-materials
The files are the evidence.
A project’s materials — PDFs, spreadsheets, notes, figures — are read through this server, then judged: does this cover the claim, is there evidence, does the context hold. The files stay where they are.
List and readNo import step and no copy into a database; the folder is the source.
Judgments, not summariesjev judgments ask whether the material supports something, not what it says.
Page-scopedMaterials belong to a project, and the judgments are about that project’s claims.
quanta-materials · four files, four verdictssimulated
PDFchapter-3-method.pdfselected
p. 41 — the sample was recruited through two channels, and the second one is not described.
coverage okevidence · 2 sourcescontext · ch. 3
XLSXsurvey-2026-09.xlsxselected
1,204 rows · 26 columns · three columns are 61% empty.
coverage okevidence · 1 sourcecontext · appendix
MDinterview-notes.mdselected
Four interviews; the quotations match the transcripts, the summary does not.
gap · summary onlyevidence · quotescontext · p. 2
PNGfigure-3.pngselected
No caption file. The axis label says n = 1,204, the text says 1,024.
gap · caption missingevidence · imagecontext · ch. 5
The judgment is the useful part; the summary you could write yourself. Each verdict points back at a page or a row.
MCP-09quanta-pages
One page reads another.
Each page in Quanta produces something: a plan, a chart, a crawl log, a sentiment read. quanta-pages lets a turn read what another page produced, so work done in one room can be used in the next.
Read, not driveIt reads another page’s output; the page that produced it stays in charge of it.
Outputs are datedAn output knows when it was written, which is how a turn tells fresh from stale.
One working folderPages share files, so reading an output is not a copy of it.
Three steps, one owner each. Step two is blocked on the API key; the crawler page already has a token that can be reused.
Failure clustering — Sep
analyse · yesterday 21:40 · 4 clusters
Timeouts cluster around two hosts. The third cluster is a cookie wall, not a network fault.
Run log · 128 new
crawler · yesterday 23:58 · partial
Two platforms complete, Bilibili stopped at page 3 with 412. Cookies are valid until Friday.
Negative share, week 41
sentiment · today 07:00 · 11%
Down from 18%. Most of the drop is one thread cooling off, not a change in tone.
Four pages, four outputs. The last turn in a project can read every room it has already been in.
Running
Started per session, spoken to in lines.
Quanta spawns a server, sends JSON on stdin, reads JSON on stdout. There is no port, no handshake to configure and nothing left running after the session ends.
The turnagent
Tool callname · arguments
Permissionasked in the window
Resulttext · image · json
stdio
One linejson-rpc
stdin{"jsonrpc":"2.0"}
stdoutJSON only — no prints
stderrlogs, freely
spawn
Python 3.10+
ask_tools.pyone file per server
Quanta configdata dir · paths
QUANTA_DATA_DIRcua also wants CUA_ROOT
Nothing to installStandard library only, except the CUA driver.
One process eachA crashing server takes down its own tool list, not the turn.
Same files as the windowMemory, schedules and page outputs are read from the app's data directory.
Install
Unzip beside the app, register, done.
01Unzip into the mcp folder.One folder per server, next to the Quanta checkout — the imports expect the app's config module.
02Run it once by hand.python3 ask_tools.py. If it starts and waits on stdin, the pipe is fine.
03Register it in Quanta.Add the command to the MCP list; the server's tools join the turn's tool list.
$ unzip quanta-ask.zip -d mcp/
$ python3 mcp/quanta-ask/ask_tools.py
# waiting on stdin — one JSON object per line
$ export QUANTA_DATA_DIR="$HOME/Library/Application Support/Quanta"
$ export QUANTA_CUA_ROOT="$HOME/.quanta/cua" # cua only
# register (Quanta → Settings → MCP)
{ "command": "python3", "args": ["mcp/quanta-ask/ask_tools.py"] }
The CUA server is the one with a real dependency: @trycua/cua-driver. Without it, quanta-screen is the fallback — a screenshot and nothing more.
Back to the app
Seven pages, and everything above plugging into them.