Your first run with Amatelier¶
Tutorial — follow every step in order. By the end you will have produced a digest file from a working roundtable and know where to find it. If you already know the tool and want task-specific instructions, see the guides.
A roundtable is one debate cycle. Ten persona agents (workers, a Judge, a Therapist, and Admin) exchange turns in a SQLite-backed chat room on a topic you define. The runner scores the turns, writes a digest, and exits. That is what you are going to produce.
Prerequisites¶
- Python 3.10 or newer
- A terminal
- One LLM credential (covered in Step 2)
Step 1 — Install¶
Install the package from PyPI:
Confirm the CLI is on your PATH:
Expected output:
If you see command not found, read the troubleshooting guide.
Step 2 — Pick a backend¶
Amatelier auto-detects three backends. Run the diagnostic:
Expected output (before you set any credentials):
amatelier 0.2.0
LLM backend
active mode: none
Available backends:
[ ] claude-code (claude binary on PATH)
[ ] anthropic-sdk (ANTHROPIC_API_KEY env var)
[ ] openai-compat (OPENAI_API_KEY or OPENROUTER_API_KEY env var)
Credentials seen in environment:
[ ] CLAUDE_ON_PATH
[ ] ANTHROPIC_API_KEY
[ ] OPENAI_API_KEY
[ ] OPENROUTER_API_KEY
[ ] GEMINI_API_KEY
...
!! No backend available. Set up one of:
- Install Claude Code (https://claude.com/claude-code)
- export ANTHROPIC_API_KEY=... (https://console.anthropic.com)
- export OPENAI_API_KEY=... (https://platform.openai.com)
- export OPENROUTER_API_KEY=... (https://openrouter.ai)
Set one credential. For this tutorial use the Anthropic SDK path — it is the shortest route:
On Windows PowerShell:
Optionally enable Naomi (the Gemini-powered cross-model worker):
Run amatelier config again. Expected:
LLM backend
active mode: anthropic-sdk
Available backends:
[ ] claude-code (claude binary on PATH)
[OK] anthropic-sdk (ANTHROPIC_API_KEY env var)
[ ] openai-compat (OPENAI_API_KEY or OPENROUTER_API_KEY env var)
Step 3 — Create a briefing¶
A briefing is a markdown file telling the agents what to debate. The shipped examples/briefings/hello-world.md exists in the git repo, not in the pip wheel. Write your own inline.
Create briefing-hello.md in your current directory:
cat > briefing-hello.md <<'EOF'
# Briefing: hello world
## Objective
Exchange one round on "what is the most undervalued idea in open-source software?" Each worker introduces themselves and offers one substantive opinion.
## Constraints
- Messages under 150 words
- One round only
- One opinion per worker
## Success criteria
A digest file with all workers present, Judge scores, and a summary.
EOF
On Windows PowerShell, create the file with your editor of choice and paste the same contents.
Step 4 — Run the roundtable¶
Execute:
amatelier roundtable \
--topic "hello" \
--briefing briefing-hello.md \
--max-rounds 1 \
--budget 1 \
--summary
Expected output (abbreviated — the runner prints per-turn lines):
[runner] topic=hello briefing=briefing-hello.md budget=1
[runner] spawning workers: elena, marcus, clare, simon, naomi
[elena] ...
[marcus] ...
[judge] scoring round 1...
[runner] digest written: <path>/digest-rt-XXXX.json
Typical cost on the Anthropic SDK with the default Sonnet+Haiku mix: $0.30 to $0.80.
Step 5 — Find the digest¶
The digest landed in your user data directory. Ask the CLI where that is:
Look under Paths::
Paths:
bundled_assets_dir ...\site-packages\amatelier
bundled_docs_dir ...\site-packages\amatelier\docs
user_data_dir C:\Users\<you>\AppData\Local\amatelier
user_db_path C:\Users\<you>\AppData\Local\amatelier\roundtable-server\roundtable.db
Open the most recent digest file in user_data_dir/roundtable-server/:
ls "$(amatelier config --json | python -c 'import json,sys;print(json.load(sys.stdin)["paths"]["user_data_dir"])')/roundtable-server" | grep digest
The digest is JSON. Open it and read the summary, scores, and turns keys. That file is the canonical artifact of a roundtable.
What you have¶
- Amatelier installed from PyPI
- An active backend
- One completed roundtable
- A digest file on disk describing the turns, scores, and summary
See a real example¶
A fully captured roundtable lives at examples/sessions/2026-04-18-self-host-vs-api/. Four SVG screenshots, a 54-message transcript, the structured digest, and a README explaining what to look for — including a real Judge GATE awarded to marcus for reframing the self-host decision.
Next steps¶
- Configure a different backend — OpenAI, OpenRouter, local Ollama
- Architecture — how the runner, Judge, and Therapist fit together
- Spark economy — how scores turn into agent currency
- CLI reference — every flag