Ollama

Ollama 0.35 and later serve System One on your own machine. ThinkThen asks them with --backend ollama, with no key and no hosted service.

Backend name	`ollama`
Address	`http://localhost:11434/v1`
Key variable	`OLLAMA_API_KEY`, not needed on this machine
Default model	`nimble`. Name `tev1` with `--model tev1`.

Set it up

Install Ollama 0.35 or later from ollama.com.
Run ollama pull nimble, and ollama pull tev1 for the smaller model.
Ask with --backend ollama. On this machine ThinkThen sends no key.

The examples below run Ollama on a second port, 11535, beside another Ollama on the same machine. So each one names the port with --url. On the default port, leave it out. A local Ollama needs no key at any port.

Descriptions travel as text

Ollama refuses a description written as an object. So this backend alone sends each description as its text. An object's what goes, and its other fields stay behind. The plan says so on its first line.

The plan names the second port and nimble. The first line says descriptions travel as text.

thinkthen check \
  --backend ollama \
  --url http://localhost:11535/v1 \
  --plan |
head -n 4

Output

thinkthen: backend `ollama` sends each description object as its `what` text, a temporary workaround for an Ollama bug, so its other fields are left out
url http://localhost:11535/v1/systemone
provider systemone
model asked unspecified
model sent nimble

exit 0

On 2026-09-30, thinkthen check on build e8f2804fe asked nimble and exited 0, with 0 critical and 3 warning findings.

On 2026-09-30, thinkthen check on build e8f2804fe asked tev1 and exited 0, with 0 critical and 3 warning findings.

Each warning names a description sent as text. Ollama 0.35 ran on the second port for these checks.

Ask nimble

The nimble model answers yes, as Jev does.

question="Does the customer ask for a refund?"

printf '%s\n' "I want to send this back." |
thinkthen decide "$question" \
  --backend ollama \
  --url http://localhost:11535/v1

Output

true

exit 0: yes

One request carries all three lines. The nimble model gives the same three answers as Jev.

question="Does the customer ask for a refund?"

cat <<'EOF' |
Please refund my order. It arrived broken.
Thanks for the quick help yesterday!
I want to send this back.
EOF
thinkthen decide "$question" \
  --lines \
  --backend ollama \
  --url http://localhost:11535/v1

Output

{"input":"Please refund my order. It arrived broken.","value":true}
{"input":"Thanks for the quick help yesterday!","value":false}
{"input":"I want to send this back.","value":true}

exit 0: the run finished

Ask tev1

The smaller tev1 model also answers yes.

question="Does the customer ask for a refund?"

printf '%s\n' "I want to send this back." |
thinkthen decide "$question" \
  --backend ollama \
  --url http://localhost:11535/v1 \
  --model tev1

Output

true

exit 0: yes

A pulled model can change under the same name. After a new pull, pass --refresh-cache once. It asks every question again and replaces the answers in the answer cache.

From your language

Each page shows how its language names a backend.

All backends · Configuration

On GitHub: github.com/botassembly/thinkthen