In het Nederlands
Interactive presentation · click anything for an explanation
A different take on MCP
Let AI code,
not call
MCP connects AI to your systems. The question is who does the work: the model that calls at every step, or a program the model writes.
Speaker notes
This is a different take on MCP. Everyone is connecting AI to their own systems now: CRM, email, accounting. Almost always in the same way.
Anthropic and Cloudflare both show it can be smarter, and cheaper. That is what this is about.
Three words to hold on to: call, code, contain.
Give it a workspace. The connection to your systems (MCP) stays exactly where it is. What changes is who does the running back and forth: the model itself, or a small program the model writes.
Speaker notes
If you remember one sentence: don't use your model as a for loop. A for loop is a bit of code that repeats something. That is exactly what we make the model do today, and it is slow and expensive at it.
Worth saying straight away: MCP, the plug between AI and your systems, is not going anywhere. This is about how you use it.
The task: post one video to Instagramclick for each round
Every step goes back through the model
The model asks one tool for something and waits. The full result lands in the conversation, and then the model rereads everything to pick the next step.
The 6,000-token transcript comes past again every round, while the model never needed to see it. A long list of tools fills the model's memory before the question even starts.
The model has become the for loop.
Speaker notes
Take a familiar task: clean up a raw video and post it to Instagram. The agent has every tool it needs.
Click through the rounds. Each round is a phone call: the model asks for one thing, gets the whole answer back into the conversation, and then rereads everything to decide what comes next. Checking the status three times is three phone calls.
Watch the meter. It does not grow evenly, because every round the model also rereads everything from the rounds before. Twelve rounds, over 190,000 tokens. For one video.
The same task, now as one programclick to follow
The workspace runs it · without the model
One program, one answer back
The model does not make the calls itself. It writes one small program that calls the tools one after another, with a loop for the waiting.
In a variable in the workspace. The model never has to read it. Only what matters comes back.
Models have read millions of lines of real code, and very few tool calls. Writing code comes more naturally to them.
Click the script or the workspace for an explanation.
Speaker notes
Same task, same tools. But now the model does not ask twelve times. It writes one program. Read it out loud: transcribe, cut, caption, upload, wait until it is ready, publish, clean up, return the link.
That program runs in a closed workspace. The transcript stays there in a variable. The waiting is a loop in the code, not three phone calls. And the only thing that comes back to the model is the link.
Same scale on the meter: from 190,000 to about 17,000 tokens. One round instead of twelve.
Three measurements, the same order of magnitude
Tokens the model has to read: separate calls versus one program. These are the makers' own measurements. The point is the order of magnitude, not the exact number.
Speaker notes
Three sources, the same picture three times. The reel we just saw. Anthropic, moving a transcript from Google Drive into a CRM: from 150,000 to 2,000 tokens. And Cloudflare, offering its whole API of over 2,500 endpoints with just two tools, for about 1,000 tokens instead of 1.17 million.
To be fair: these are the makers' own measurements, on tasks they chose. Cloudflare also gives an in-between figure of 244,000 if you count more cleverly. The order of magnitude holds.
Code that runs on its own only gets the keys you hand over
The workspace
The script runs apart, not on your own systems.
- Limits on time and memory
- Only the connections you allow
- Every run logged
The rights
Read first. Write only with your approval.
- Own key per connection
- Start read-only
- Never messages clients on its own
The data
What the model does not need to see, it does not see.
- Intermediate results stay inside
- Names replaced by codes
- The model only sees the outcome
One loop can make a hundred changes. That is why the workspace starts read-only.
Speaker notes
Here is the flip side. With separate calls you see every step and can stop every step. With a program you approve once, and a loop can change a hundred things. So the workspace has to be closed.
Three things. Isolated: the script runs apart, with limits on time and memory. Read only: start with a key that cannot change anything. And privacy, which is a bonus: because intermediate results stay in the workspace, you can replace names and personal data with codes. The model never sees them.
- No: MCP stays the plugCode Mode uses the same connections. A program calls them instead of the model.
- Calling is fine for short questionsOne or two steps and small answers: separate calls are simpler and easier to check.
- Coding has a priceA secure workspace with limits and monitoring, and a model that writes good code. Anthropic says so itself: it brings its own complexity.
Speaker notes
On social media people are shouting "RIP MCP". That is not right. MCP is the plug, and Code Mode simply plugs into it. The difference is who uses the plug.
For a short question, "what is the status of order 1234", calling is fine. Coding pays off with sequences, repetition and large amounts of data. And coding has a cost: you need a secure workspace, and the model has to write the code well.
Five things we take into account
- CodeWe choose calling or coding per taskA single question can go to the model directly. A fixed sequence with many steps or a lot of data goes into one script, so the model only sees the outcome.
- ContainA connection never runs as administratorIt runs on its own user and cannot reach the system settings. What it does not need, it cannot touch.
- ContainReading and writing are separatedSeparate keys for reading and for writing, decided per person. Keys never travel through a chat or an email.
- ContainPersonal data stays where it belongsNo care or personal data of clients in the knowledge base, with a scan beforehand. Nothing goes to a client without a person approving it.
- ContainEvery change can be tracedWhat a connection changes in Odoo is in the chatter of that record. What happens in the knowledge base is there with a date.
Speaker notes
These are the choices we make when we connect AI to your systems. The first is about coding: per task we decide whether the model may make the calls itself, or whether it becomes a script. The other four are about containing.
A connection never runs as administrator. Reading and writing are separated, with their own keys. Personal data stays out of the knowledge base, and nothing goes to a client without a person approving it. And what changes can be traced: in Odoo it is in the chatter.
Call where you can, code where it pays off
| Situation | How we do it |
|---|---|
| A single question to Odoo | CallThe model uses the connection directly. Fast, and every step can be checked. |
| Collecting news, transcribing, publishing | CodeOne script runs the whole sequence. The model only sees the outcome. |
| An Odoo connection in Code Mode form | Still openWe do not have one yet. A first trial would be read-only, on our own database. |
Speaker notes
We work with both every day, and this is where we stand. A single question to Odoo: calling. A fixed sequence, like collecting news, transcribing and publishing: one script does it, and the model only sees what comes out.
And we are honest about our own gap. We do not have an Odoo connection in Code Mode form yet. When we build one, it starts read-only.
Call, code, contain
Three images
- CallTen phone calls, rereading the whole file after each one.
- CodeOne work order for an assistant in a closed room.
- ContainThat room only has the keys you hand over.
Three moments to recall it
- Day 1Explain it to one colleague in three sentences.
- Day 3When is calling fine, and when does coding pay off? Answer it without this deck.
- Day 10Hold one AI connection you use up against the five points. What is missing?
Speaker notes
Three images to hold on to. Call: ten phone calls, rereading the whole file after every call. Code: one work order for an assistant in a closed room. Contain: that room only has the keys you give it.
And recall it, instead of rereading it. Tomorrow, explain it to someone; in three days, the recall question; in ten days, hold one connection up against the five points.
not call.
Our advice: make sure an AI task that keeps coming back does not go past the model at every step. Let it code, and contain it from day one.
Speaker notes
Back to the sentence from the start: let AI code, not call. And the three words: call, code, contain.
Our advice to take home: make sure an AI task that keeps coming back does not go past the model at every step. That saves time and money. Let it code, and contain it from the start: read-only, own keys, personal data out of sight.