Agents

An AI agent that can run shell commands uses ThinkThen the way a person does: it pipes in the text, asks a question, and reads the answer or the exit code. --json gives it the details, and --check gives it a yes or no it can branch on.

echo 'I was charged twice for one invoice.' |
  thinkthen decide --json \
    'Does the customer request a refund?' |
  jq -c '{status, answer, probability}'

Output

{"status":"answered","answer":true,"probability":0.7}
Copy

The agent gets one short line: the status, the answer and the chance of yes. jq -c keeps only those fields, so the agent reads a line rather than a page.

Branch on a yes or no

With --check, the answer is also the exit code, which most shell tools hand back to the agent with the output:

echo 'Where is my parcel?' |
  thinkthen decide --check \
    'Does the customer request a refund?'

Output

false
Copy

No exits 1 here. A failed call has a code of its own, with the error on standard error. Exit codes lists every code.

Let the agent learn the command

thinkthen --help lists the ten functions, and thinkthen decide --help shows a working example, every option, the settings it reads and the exit codes. Each function's help also names the address of its page here, so the agent can read more.

When the text sits in a shell variable, --text "$message" passes it without a pipe. Many messages and files go through a pipe or file names, as File types shows.

Read these pages as Markdown

Every page on this site is also served as Markdown: drop the last / of its address and add .md. This page is at https://thinkthen.dev/command-line/agents.md. https://thinkthen.dev/llms.txt lists every page with one line on each, and https://thinkthen.dev/llms-full.txt holds every page in one file.

To set an agent up, add two lines to the instructions it reads at the start of each task:

Use the thinkthen command for questions about text.
Its docs are at https://thinkthen.dev/llms.txt.

No MCP server

ThinkThen has no MCP server. An agent with a shell tool needs none: the command is the whole interface, and its help and these pages tell the agent how to call it.

Next

decide shows each answer a yes-or-no question can give.