poindexter

The open-source AI content pipeline from Glad Labs, as one installable package: the poindexter command-line client, the service layer it delegates to, the plugin registry, and the standalone brain watchdog daemon. The CLI, the worker, and the MCP servers all import the same poindexter.* modules — one schema, one client, zero drift.

Install

pip install poindexter
poindexter --help

Python 3.13. The package talks to a PostgreSQL database with the pgvector extension; poindexter setup writes ~/.poindexter/bootstrap.toml and runs the schema migrations against it.

The bare install is the CLI and the client library. The generation stacks are extras, so a machine that only operates the pipeline does not download them:

pip install "poindexter[pipeline,qa,rag]"   # a full pipeline node

A command that needs an extra you did not install fails with a plain ModuleNotFoundError naming the package — never a silent degradation.

To run the whole stack (worker, brain daemon, Prefect, Grafana, ...) rather than the client alone, clone the repository and start the consumer compose file:

git clone https://github.com/Glad-Labs/poindexter
cd poindexter
docker compose -f docker-compose.consumer.yml up -d

Quick start

# 1. Point the CLI at a Postgres instance (pgvector enabled) and run migrations.
poindexter setup

# Non-interactive — takes an explicit DSN:
poindexter setup --db-url="postgresql://user:pass@host:5432/poindexter_brain"

# Re-run the sanity checks against an existing bootstrap.toml:
poindexter setup --check

# 2. Ask questions of the shared memory spine:
poindexter memory search "why gemma3"

# 3. Inspect the current pipeline task queue:
poindexter tasks list --limit 20

# 4. Read or change runtime app_settings:
poindexter settings list
poindexter settings set enable_pyroscope true

Every command respects -v / --verbose for client-side info logs and -h / --help for a per-subcommand usage summary.

Command groups

Run poindexter --help for the full list (forty-odd groups) and any group with --help for its subcommands, e.g. poindexter memory --help.

Import root

Everything ships under one package:

from poindexter.services.site_config import SiteConfig
from poindexter.plugins.registry import get_taps
from poindexter.brain.bootstrap import resolve_database_url

The pre-1046 flat spellings (import services.x) are gone — there is exactly one module object per name, and nothing is installed at the top level except poindexter.

Configuration

poindexter reads its infrastructure secrets (mostly: a Postgres DSN) from ~/.poindexter/bootstrap.toml, which poindexter setup writes for you. Everything else — API keys, feature flags, quality thresholds, model pins — is stored in the app_settings table in Postgres and can be managed with poindexter settings.

There are no required environment variables. DATABASE_URL, if present, is accepted as a convenience override.

How versions work

The PyPI version tracks the upstream poindexter repository via release-please. Each release on main cuts a tag v<major>.<minor>.<patch>; that tag fires the publish workflow, which builds the wheel, installs it into a clean venv, runs poindexter --help, and only then ships it.

Breaking changes in the CLI show up in the repo's CHANGELOG.

Development / working from a clone

If you've cloned the upstream repo you don't need PyPI — everything is already wired up:

# From repo root:
cd src/cofounder_agent
poetry install --extras "pipeline qa rag"
poetry run poindexter --help

To build the distribution yourself (the manifest is src/cofounder_agent/pyproject.toml):

cd src/cofounder_agent
uv build                     # or: python -m build
pip install dist/poindexter-*.whl
poindexter --help

License

Apache 2.0. See LICENSE in the upstream repo.

Support

• Bugs and feature requests: https://github.com/Glad-Labs/poindexter/issues
• Documentation: https://github.com/Glad-Labs/poindexter#readme
• Security: see SECURITY.md