09mes-mcpOngoing
“I don't care about a pretty interface. I want a set of tools my everyday LLM can call”
me, watching everyone build themselves a Jarvis
Not one more assistant: hands for the one I already have.
Five tool servers, called in plain language by the model I use every day: read, produce, send. 63 tools, a single way out, compartmented by project.
- From the first file to the first image out
- 1 day
- What's left of it today
- 63 tools, a single way out
- First file
- 6 June 2026
$ open mes-mcp Private servers, in daily use · designed, coded and shipped, solo
ping projet: studio9 envoi OK · projet « studio9 » · canaux d'envoi disponibles : email send channel: email · to: personne@example.org · attachments: 1 (livret.pdf) · request_id: livret.v3 Admis a l'outbox (id 7c1e...). Le worker livre en arriere-plan ; suivre via GET /v1/sends/7c1e...
What the sentence meant
In June 2026, everyone was building themselves a Jarvis, with the beautiful interface that goes with it. I did not need an interface. I needed the model I already use every day to be able to act: read, produce, send. Not one more assistant to open, hands for the one I have.
An MCP server describes each of its tools in one sentence; the model calls it when the conversation asks for it, and reads an answer written for it. The server writes nothing and decides nothing: it carries out, under guard rails. Its power is measured by the list of what it can do on a plain instruction.
Where it lives: on Instagram, among other places. Not a robot that posts: an account kept in plain language, with not a gram of automatic content, for people whose communication is not what they think about. The first of them is my little brother, a table tennis champion. The tool is growing mostly for him: so that somebody looks after his communication while he does what he loves, playing.
What was built
One repository, five MCP servers and three command-line tools. Two servers run on a rented server (envoi, image-gen); three stay on the Mac because they touch personal data or sites that block data centres (reel-extract, wa-extract, relations). The everyday model, Claude in conversation or on the command line, is the only caller.
01Extract and understand any content at all
A Reel shared from the phone lands in an iCloud queue; the tool downloads it, transcribes it, keeps the caption and the metadata, then files it under a taxonomy that never overrides a human sort. A WhatsApp thread comes in through the app's own export, never through an automation plugged into the account; the voice notes are transcribed locally, free of charge. A site is probed before being copied. All of it ends up in a database you question in plain French, with the source message or Reel and its context.
That reader never touches the account: it reads an export made from the phone, in thirty seconds. Sending is another trade, and goes through the gateway, further down.
Tools, as they are called
- extract_reel
- drain_inbox
- search_reels
- classify_reel
- apply_human_sort
- prepare_frames
- search_messages
- get_message
- export_days
- transcribe_pending
- ditto probe
02Reproduce the code of a site, or of a single element
ditto copies what it is asked for, coldly: it probes the page, photographs it at every width, films its motion, reproduces a single element as a standalone file with its states, clones the whole site, then measures in pixels what separates the copy from the original.
03Create images and videos, with a real workshop
A workshop for images and videos: several providers behind a single passage, a gallery, an approval before every euro. The caps live server side: they are refusals, not instructions.
It is a Higgsfield or a Midjourney without the subscription, and without leaving the conversation: I describe, it prices, I approve, the image arrives. I pay for the models and nothing else.
One case: the bee that crosses a site
The bee flying on the Abeille Angloye site came out of here, without leaving the editor and without opening a single third-party tool. Three sentences, three steps.
-
“A bee in gouache, in profile, translucent wings, transparent background. Six variants.”
The plan is priced before anything is spent. The six arrive in the gallery, the association chooses.
-
“This one, set it flying. Four seconds, looping.”
Video costs more: it waits for a human approval in the gallery before the smallest paid call.
-
“Cut it out, transparent background, ready for the web.”
It flies today on the association's site, and on this project's page.
Tools, as they are called
- generate_image
- upload_image
- estimate_cost
- generate_video
- plan_video_batch
- run_job
- submit_choice
- manage_asset
- search_assets
- video_frames
- video_last_frame
04Produce documents, charts and animations
Two deterministic chains. pdf-gen prints the handover booklet given to a client: configuration, time statistics and git milestones, and a file identical to the byte from one machine to another. video-produit shoots the presentations: a video written in animated HTML, a voice-over, a frame-by-frame render to MP4.
Types and skills
- livret
- hyperframes
- rush-graphique
- rush-generation
- media-use
- onboarding
05Send, without leaving the conversation
A single gateway lets messages out: email with attachments, WhatsApp, text message, notification on the phone. You talk to it from the everyday model, without opening a tool: write to this person, attach the quote, schedule it for Monday morning.
Every project has its compartment: its addresses, its limits, its address book. Studio 9 cannot write with Toul'Barz's identity, nor dip into its address book.
Plugging one more in takes a few lines. Yours would be one of them.
Before it leaves, every message passes some twenty checks, closed by default. The answer fits in a word: allowed, duplicate neutralised, preview, scheduled, or refused, with the reason and the wait.
The gateway delivers, it does not write: the text always comes from the person, or from the model they are driving. And a send to dozens of people never goes out on its own: somebody reads one copy, then clicks.
Tools, as they are called
- send
- identities
- limits
- list_sends
- cancel_send
- whatsapp_health
- campaigns
- create_upload_url
- docs
- ping
06Keep the memory of relationships
One record per person, in plain text, readable without the tool. The model writes an exchange into it, updates what is known, brings back the due dates.
Tools, as they are called
- fiche
- chercher
- echeances
- consigner
- maj_personne
- fusionner_personnes
- file_de_travail
- verifier
07Instagram, what comes next, in order
Today the tool reads Instagram: shared Reels become a bank of ideas. Next, posting on the account. Then answering messages: the model proposes, a person approves, off it goes.
What mes-mcp does not do
- Nothing goes out by itself: the person asks, through the model they are talking to. The tool carries it out, within the limits set for it.
- It does not post on Instagram yet, does not comment, does not send direct messages: reading exists, writing is the next brick.
- It does not run unsupervised: the sending circuit breaker suspends a project that runs away until a human resets it.
The result
“I don't care about a pretty interface” became: 63 tools callable in plain language, on 5 servers, behind a single way out, compartmented by project. The first tool was plugged into my everyday model on the day of the first file; the first real image came out of the server the next day; the first message went through the door on 24 June 2026.