agtcli

agtcli

Build CLIs an agent can actually drive. A runtime you import, a scaffold that generates a working tool, and a checklist that runs.

The command, the module, and the PyPI distribution are all agtcli.

The rules and the incident behind each are in PLAYBOOK.md. This file is how to use the thing.

Why a library and not a template

The same contract copied into five projects is five things that drift. orly and mercadona-cli independently grew the same exit-code taxonomy, the same compact/--raw split, and the same atomic-write helper, in two languages, and the copies already disagree.

So the mechanical half is imported and tested once. agtcli check covers the half that is still yours.

Install

pip install agtcli     # or: uv tool install agtcli
agtcli --help

From a clone instead, which is what you want if you are changing the contract:

uv sync && uv tool install --editable .

Either works, and the difference shows up in what you scaffold. A generated project depends on agtcli, and agtcli new pins that dependency at wherever this copy came from: nothing to pin from an index, the local path from a clone, the same commit from a git install. So a project scaffolded by the released agtcli builds anywhere; one scaffolded from your clone builds where that clone is.

Recipes

Scaffold a CLI and confirm it works

agtcli new hansard --description "Search UK Hansard debates." \
                    --triggers "hansard, what did parliament say about X"
cd hansard && uv sync && uv run pytest -q     # 14 passing
uv tool install --editable . && hansard --help

You get a tool that already has typed exit codes, a compact default with a --raw escape hatch, atomic 0600 state, a credential-rotation hook, a blast-radius guard, a thin skill, and contract tests that pass. Then replace src/hansard/api.py with the real service.

Audit one before shipping

agtcli check ./hansard          # exits non-zero while anything fails
agtcli check ./hansard --json   # one object per check
agtcli rules                    # just the checklist

Gate a commit or a CI job on the checklist

agtcli check . --json | jq -e '.failed == 0'

Find which rule a check is about

agtcli rules --json | jq -r '.[] | select(.name=="truncation-marked") | .why'

Fifteen items are checked by reading the project. Nine cannot be, so they are reported as unverified rather than counted as passes.

$ agtcli check ./hansard
  ✓ skill-shipped: skill/SKILL.md
  ✓ exit-codes-typed: inherited from agtcli.build
  ✓ truncation-marked: cuts are marked
  ✓ blast-radius-capped: costly commands consult the guard
  ...
  ? dumb-subagent-test-run: the only step that reliably finds what you missed

9 item(s) cannot be checked mechanically; verify them yourself.

What you import

import agtcli
from agtcli import Session

app = agtcli.build("mytool", help="...", notes=AGENT_NOTES, unit="EUR")

@app.command()
def get(ctx: typer.Context, id: str):
    s = Session.get(ctx)
    s.out.emit(record, compact=lambda r: {"id": r["id"]})

@app.command()
@agtcli.costly                       # cannot return without checking the cap
def submit(ctx: typer.Context, id: str):
    s = Session.get(ctx)
    s.guard.check(total, f"submitting {id}")

main = agtcli.main_for(app)

Errors are typed and carry what to do next:

$ hansard submit 123 --json ; echo "exit $?"
{
  "error": "refused",
  "message": "submitting 123 is irreversible and --yes was not given",
  "exit_code": 6,
  "retryable": false,
  "remedy": "re-run with --yes once a human has agreed"
}
exit 6

The skill

skill/ is meant to be symlinked, so it versions with the code:

ln -s "$PWD/skill" ~/.claude/skills/agtcli

A CLI is invoked, not found, so the skill body is a one-liner that points at --help. What the agent discovers progressively is the tool: tool --help for the verbs, rules and exit codes; tool <verb> --help for one command, with the global flags carried down so --max-cost is still visible on the command that spends; and a next-step hint in the output itself.

Development

uv run pytest -q

Every test is a rule from the playbook or a bug that actually happened. If you change the contract, a test should fail.