← All workshops · Slides of workshop 4

Workshop 4 - Plug in tools: facilitator sheet

For the person who runs the workshop. The participants' page is the README.md of the repository of the workshop, yal-workshop-tools (source: workshops/4-tools/ of the facilitator's repository), and their one-page sheet sheet.md.

⚠️ Important. In mode D, the model can read every file of the repository, including its README.md, and replay does not flag it. The participants' pages never give the expected answer of question 8, nor which status to choose for Antwerp. Keep it that way when you adapt them, and keep this sheet out of the repository.

Question of the day: what happens when the chatbot acts outside? Message to leave with: a tool brings what no document can hold, today's facts; it runs with your rights and may send data away, so you read it before you plug it in; and its data is only as good as the person who fills it in, from the official source, on the day. Two points to bring out in the discussion: tools combined cross data in seconds (end of Part 3: purpose principle of the GDPR and of the Informatique et Libertés law), and the guardrails are shown, not set, by the participants (end of Part 4). With young people: read-only tools, no personal data, public sources only.


Before the workshop (about 1 hour)

Time What How you check it
15 min One Anthropic API key per pair, all in one workspace for the room, capped at $3 × the number of pairs (see Who pays). workshop-login says Key OK
10 min Read the official sources of the day, the morning of the workshop: the page of each territory, given in the source of its entry in mcp/fire-danger.json, and facilitator/LIVE-SOURCES.md. Note for each one: is there a level today, or a sentence saying the season is over, or nothing? A pair from another territory needs the address of its official page (the page asks for it in "Before you start"): remind them at the end of workshop 3, and look for it with them if they forgot. you know what each pair should write
20 min Your own run: a new Codespace from https://codespaces.new/Youth-AI-Lab/yal-workshop-tools, workshop-start 4, workshop D, /mcp, question 8 before and after you fill in Antwerp, the Pla Alfa question twice (with and without "Use the time tool:"), WORKSHOP_WEATHER=1 workshop D and the two weather questions, replay D 8; openwebui start in a second terminal at the start, then openwebui chatbot --name "…" and question 8 in Open WebUI each step as the page says; Live tool ready: fire-danger (fire_danger, territories), read-only.
10 min Prepare the page with a hidden sentence for Part 4: save trap-page.html on your computer (right-click on the link). Never in a pair's Codespace, never before a replay. the page opens, the hidden line is there
5 min Print the one-page sheet (one per pair; translate it first if your room needs another language: keep the commands and what the terminal prints in English) and the exit cards (one per person: the three questions of Exit card) -

How to create a key and set a limit: Open WebUI guide, Your Anthropic API key.

Or OpenRouter keys (sk-or-): same commands, no Claude model, and the whole room on one model so that totals compare (OpenRouter guide). The figures of this sheet come from Claude and do not hold for OpenRouter (its measures). Cost per pair with the default model, DeepSeek V4 Flash: about $0.01 for the terminal commands of this workshop and about $0.12 for workshops 1 to 5, played once on 2026-11-05 in a container of the image 2026.11.4 (OpenCode 2.1.4), not in a Codespace and not with a group (measures). Not counted: replay RAG and the Open WebUI chatbots. A pair that asks more questions spends more: take it as a minimum. The workshops are for adults: no young person types in the Codespace or holds a key, and OpenRouter asks for users of 18 or older. Young people only use a chatbot in Open WebUI, in their browser. When what they wrote goes through OpenRouter, leave out names and personal details, and type workshop-login --europe first (SETUP.md, Young people and OpenRouter; OpenRouter guide, Part 8).

ℹ️ Note. Who pays. Each partner organisation pays for its own sessions: its own Anthropic account, workspace and keys, on its project budget. One workspace for the room, capped at $3 × the number of pairs: 8 pairs, $24. At Antwerp, the most advanced pairs used less than $2 in a 2-hour workshop. The limit is on the workspace, not on each key, so one pair can use up the share of another: keep the Anthropic console open during the workshop and watch the spending. Not checked: if the console limit is per month, the workshops of the same month share it. Then count $3 × pairs × workshops of the month, or raise the limit before each workshop. Indicative cost per pair for the whole series: TO BE COMPLETED at the rehearsal; this workshop's measured cost is under "Cost per pair".


Timing

Time Part What you do What the room should have at the end
0:00 1. Welcome (20) Setup (SETUP.md, steps 1 to 5), workshop-start 4 --from my-work.zip: the newest zip (maybe my-work (1).zip, its exact name between quotes), check the saved on date. The first workshop D asks again to trust the folder: No, exit is proposed. openwebui start in terminal 2 (a few minutes: the pairs go on without waiting). Ready for workshop 4; openwebui start running in terminal 2
0:20 2. Our tool (20) workshop D, /mcp, question 8; fill in Antwerp together on the big screen from its official page, then each pair in its file; question 8 again; make everyone press Ctrl+O. Antwerp filled in, dated today, with its source
0:40 3. Servers that already exist (30) .mcp.json; the Pla Alfa question twice; the weather; the gauge. each pair can say what each server reads and sends away
1:10 4. Read before you install (25) fire_danger_server.py and the five questions; your demo of the hidden sentence (the result of our test stays folded until the discussion); each pair fills in its own territory, then asks for a change of mcp/fire-danger.json and answers 3. No to Do you want to make this edit…? (not timed). one entry per pair with its source; each pair has seen a permission prompt
1:35 5. Closing (15) Prediction; replay D, score D 8 11 12 (no question 12: score D 8 11); compare the rows with column C, not the Total (no C score for 8: score C 8); openwebui chatbot --name "…" and question 8 in Open WebUI (View Result from fire-danger_fire_danger); workshop-save (bring the newest zip), exit cards. Short of time: show your own chatbot on the big screen instead of the pairs' (timing with a group: not measured). column D for 8, 11, 12; my-work.zip downloaded
1:50 Margin (10) - -

The data of the day (out of the participants' pages)

What to write for Antwerp. Read https://natuurenbos.be/waarschuwingen the morning of the workshop.

Expected answer to question 8 on the day (for you and for the judge, which reads it from the file): the level the pairs wrote, with its source; or, with season-over or no-level-shown, "no official level today", with the sentence of the source. A level that is not in the file is wrong (0). In the other modes, "I don't know" = 1.

Territories without a daily public level. Naples: no public daily page found (status no-level-shown, scale and source still TO BE COMPLETED in the file). La Rochelle and Aix-en-Provence, out of season: the prefecture maps publish no data. For La Rochelle, the order 26EB142 says what applies then: "En cas d'absence de données sur le niveau de vigilance, seules les dispositions générales prévues par la réglementation en vigueur s'appliquent." (corpus/text/30), a good source_says. Barcelona: the Pla Alfa runs all year.

A recorded day, to go further (optional). FIRE_DANGER_RECORDED=1 workshop D adds a third tool, recorded_example, that reads mcp/fire-danger.recorded.json: a past day, always marked "RECORDED ON <date>, NOT TODAY". The file holds no real day yet: TO BE COMPLETED, copied by hand from the official source with its evidence (method: facilitator/ATELIER-4-MCP.md, section 2.5). Do not use it with the pairs until it is filled in.


What we measured

All on image 2026.10.34, 2026-10-30, in a throwaway container of the image, on a copy of the repository, claude-sonnet-5-5, Claude Code 2.1.303; with a test key capped at $2 (about $0.5 spent). Details: facilitator/ATELIER-4-MCP.md, section 6.

What Measure
workshop-start 4, new repository Feurisson, the Benevento document added ("The Benevento document was not in the corpus (workshop 3): it is now …"), First command: workshop D
workshop-start 4 --from my-work.zip system prompt, skill, Benevento and index lines restored; "nothing needed"
/mcp in mode D 2 servers, ✔ fire-danger 2 tools, ✔ time 2 tools; with WORKSHOP_WEATHER=1: 3 servers, ✔ open-meteo 17 tools (the model gets 2)
The tool calls on screen a grey line Called fire-danger 2 times; Ctrl+O shows ● fire-danger - fire_danger (MCP)(territory: "Antwerp") and the JSON sent back
Question 8, file not filled in fire_danger, territories; no level, reason and source given; 8,000 tokens; 6 s; $0.03
Question 8, Antwerp season-over "…no official fire danger level for Antwerp today", the ANB sentence quoted and translated; $0.007
Bouches-du-Rhône summer rules today time not called: the date came from the field today of the fire danger tool; right answer (out of the 1 June - 30 September period) from the order
"What time is it now in Naples?" Feurisson refuses: out of its topic
Pla Alfa question, plain time not called; "I don't know" (the update hours found, not the time now)
Pla Alfa question, "Use the time tool: …" time called (Europe/Madrid), right answer; 7 s; $0.015
Weather: wind tomorrow in La Rochelle weather_forecast with coordinates (no geocoding); "This is a weather forecast… It is not a fire danger level."; about 14,000 tokens. Weather plugged in, not called, fire danger tool only: 13,868 against 7,845 without the weather, so about 6,000 for the server (no fill-in state noted for the 13,868 run)
Weather: "Given the weather, what is the fire danger level…" the fire danger tool, no level, "I didn't use the weather to guess a level" (one try only)
La Rochelle no-level-shown with the sentence of the order the sentence quoted and translated, "No level does not mean no risk"; $0.011
replay D (11 questions, no question 12) 1 min 54, $0.24; judge: question 8 = 2, question 11 = 2
OpenCode, replay D 8 same answer, same two tools; 5 s; $0.02
The page with a hidden sentence (Claude Code and OpenCode, mode D, image 2026.10.33) neither followed the hidden sentence; OpenCode pointed it out, Claude Code did not (facilitator/injection-demo/README.md)

Cost per pair: about $0.45 for what the page asks (replay D $0.24, question 8 twice and the other questions about $0.20, the judge $0.0005). Each pair's $3 share of the room's limit leaves a margin.

The tool in Open WebUI (Part 5, image 2026.11.4). From workshop 4 (the kit has mcp/ and .mcp.json declares fire-danger), openwebui start also starts the fire-danger server, in its own container: the Codespace's user, read-only file system, no capabilities, no published port, no key, mcp/ mounted read-only (fire-danger.json read again at each call), on an internal Docker network shared with Open WebUI only, with a random token renewed at each new Open WebUI container. Only fire_danger and territories are served (not recorded_example, even with FIRE_DANGER_RECORDED=1; not time; never Open-Meteo). openwebui chatbot gives the tool to the pair's main chatbot, never to the --rag one; chatbot-export does not carry it. The class mode of the Open WebUI guide has it from image 2026.11.8 (not published yet), with other rules: see "Reuse with young people" below. Measured on 2026-11-04 with a fake model and a fake key, no route out (facilitator/openwebui-prototypes/next-image/): the two tools offered by their full names, level: null and why_no_level with the kit's undated file, the level with a file dated today and live; unlisted names "not found"; 401 without the token, 421 with a forged Host. Two limits: the Codespace itself can reach the internal network (the token is the lock there), and Open WebUI's API, without accounts, shows the token and the provider key to whoever reaches it, so port 3000 stays private. With a real model, on 2026-11-05, on a laptop, not in a Codespace (facilitator/openwebui-prototypes/real-model-local/): Claude Sonnet 5.5, one try, called territories then fire_danger(Antwerp) at the first attempt and gave no level, because the file was not dated today ($0.021, results-claude/); with qwen3:8b (Ollama), the first fire_danger call after Open WebUI started was not run in one try out of three (status incomplete), and that answer was empty: tell the pair to ask again, as the page says (results-thinkpad/). Not tried: a real Codespace (Docker in Docker, a Codespace that sleeps and resumes), the screen, replay RAG in this repository, the time openwebui start takes here.


Leading the discussions

After Part 2. "Who decided to call the tool, and who decided what it answered?" The model decided to call it (from its description), the file decided the answer, and a person filled in the file. The "Think about it": the person who filled in the file, from the official source, is responsible.

During Part 3. Two moments:

  1. The Pla Alfa question: the model did not call time until the question named it. "A tool the model never thinks of still costs its descriptions in every message." Show the gauge.
  2. The weather: what left the Codespace? The coordinates of La Rochelle, sent to Open-Meteo by a server that a volunteer wrote, running in our Codespace (Open-Meteo also sees the address of the Codespace). "For a town, fine. For the address of a young person?"

Your demo in Part 4. Drop trap-page.html in your Codespace, workshop D, ask "Summarise trap-page.html for a group of young people.", show the answer, then open the file in the editor: the hidden line appears. Discuss: who wrote it, who read it, what else could it ask? The participants' page keeps our test result folded ("After the discussion: what happened in our test"): open it with the room after the discussion. Delete the page before any replay.

After Part 4. "Would you plug in a server you cannot read?" The five questions are the habit: who wrote it, what it reads, what it sends away, does it write, is its version fixed.


Exit card

  1. Why can only a tool give today's fire danger level?
  2. What can a server do in your Codespace, and how do you check it before you plug it in?
  3. What is the difference between a weather forecast and the official fire danger level?

What a good answer contains: (1) it changes every day and is decided by an authority; no document holds it; the tool reads what a person copied from the source, on the day; (2) it runs with your rights: it can read files and send what it gets to a service; read its code and its options, or ask someone who can; (3) the forecast is a measure of the weather; the level is a decision of an authority, with its own scale and rules.


Fallbacks

Problem What to do
fire-danger.json is malformed (…) The line and column are in the message: a missing comma or quote. Fix it together on the big screen.
The tool says the file is not dated today "date" at the top of the file, year-month-day, today.
A territory has no daily public level no-level-shown or season-over with the sentence of the source (see "The data of the day"). Never another index.
Open-Meteo does not answer (network, limit) Skip the weather step; the descriptions and the gauge are still worth showing with /mcp.
/mcp shows ✘ Leave and start workshop D again.
The model does not call time Expected: see Part 3. Name the tool in the question.
A pair has no save, so no column C Compare with a neighbouring pair.
Column C has no score for question 8 score C 8: the answers of workshop 3 are in answers/ (restored from my-work.zip).
A pair restored an old save (saved on is not the end of workshop 3) workshop-start 4 --from the newest zip, its exact name between quotes.
A pair uses OpenCode Same tools, no question before a tool call; it asks before every command it wants to run in the terminal and every file it writes.
Open WebUI is not ready at the closing Skip openwebui chatbot; show your own chatbot and question 8 on the big screen (built in your own run).
! The live tool fire-danger could not start Most likely an edited mcp/fire_danger_server.py: undo the change, then openwebui start again (chatbots and documents are kept).

After the workshop

Reuse with young people. "Can we go to the forest today?" is a question young people ask: only the official source can answer, on the day, and a chatbot must say when there is no level. The chatbot with this tool is the pair's chatbot in Open WebUI (page, Part 5, Step 2, openwebui chatbot): read-only, no personal data, public sources only. It lives in the pair's own Codespace, an Open WebUI without accounts whose port stays private, and chatbot-export does not carry it. With a group, the tool reaches the young people through the Open WebUI guide (Step 5 - Add the live tool), from image 2026.11.8 (not published yet: the guide pages must not go out before it). There:

Measured on 2026-11-06, once, without network, Open WebUI 0.12.5, a fake model and a fake key, the kit's openwebui on the host Docker (facilitator/openwebui-prototypes/class-tool/): a participant account got the tool and the level from a file dated that day, level: null from the template; without the grant, no tool at all for the participant while the admin still got the level; the warnings for a bad file, a file of the day before, a link and a tracked file; no route out for the server; workshop mode unchanged. Not tried: the image itself (not built), a browser, a real model, a real Codespace, a group, share close.

Another way, for adults, without Claude Code, never with young people: OpenCode guide, Part 6 - Tools with MCP and What the rules cannot stop.