Praxis Agents OS
View on GitHubRead the docs

Put AI agents to work in your business

Praxis Agents OS is your platform to create agents, connect them to the tools your team already use, schedule their work, approve the actions that matter, and see exactly what happened afterwards.

View on GitHubWatch the demo

Open source. Runs on your own infrastructure with Docker. Bring an OpenAI, Anthropic or Google key, or point it at an open-source model.

Waiting for approvalA Gmail agent drafts a message, then stops. Nothing is sent until a person approves it. Recorded in a demo workspace.
Connects to
  • Gmail
  • Google Ads
  • Google Analytics
  • Airtable
  • BigQuery
  • Notion
Runs on
  • OpenAI
  • Anthropic
  • Google
  • Open-source models

Ask for the work. Watch it happen.

Give an agent a task, or put it on a schedule. It plans, uses the tools you allowed, and shows every step as it goes.

A conversation showing an agent's plan and a sequence of completed Google Ads report tool calls
A Google Ads agent working through a search-term review. Each card is a real tool call with its account, its status and its result.
  1. 01

    Ask, or schedule

    Type a request in plain language, or set a cadence such as every Monday at 09:00.

  2. 02

    Plan and act

    The agent works through a visible plan, calling the tools you allowed: run a Google Ads report, search the knowledge base, read a file.

  3. 03

    Stop for approval

    Anything that sends, spends or changes an external system waits for a person. You can edit it before it goes.

  4. 04

    Keep the record

    Results, charts and documents stay in the workspace. Every tool call gets an audit entry with who asked and who approved.

The repeat work that lives in ad accounts, inboxes and spreadsheets.

  • Weekly ad account reporting

    A scheduled agent pulls Google Ads and Analytics data, writes the commentary, and leaves the report ready to review.

  • Search term clean-up

    An agent reviews search terms, proposes negative keywords, and waits for sign-off before touching the account.

  • Questions over your data

    Ask a BigQuery agent for a number. It runs a read-only query, shows the SQL, and returns a chart.

  • Drafting from the inbox

    A Gmail agent reads the thread, drafts the reply, and holds it until someone presses Approve and Send.

You decide what an agent can do on its own.

Every tool has a policy: Off, Approval, or Auto. Reads can run unattended. Writes wait for a person by default. Some, like changes to a Google Ads account, cannot be switched to Auto at all.

40
tools across six integrations
19
write tools that can never run without approval
400
days of audit log, with who asked and who approved every call
Per-tool policy switches for a Google Ads agent showing Off, Approval and Auto options
Tool access for a Google Ads agent. Negative-keyword changes are locked to Approval. The read-only report tool runs on Auto.

Agents, context, schedules, integrations. One place for all of it.

The sidebar has five items. Between them they cover what an agent needs to do useful work and what a team needs to trust it.

Click any screen to see it full size.

  • Agents

    Give each agent a name, instructions, a model and a set of tools. Choose OpenAI, Anthropic, Google or an open-source model, and a tier from light to max.

  • Context

    A knowledge base, skills, files and memories. Agents search and cite them, and you can correct what they remember.

  • Schedules

    Recurring, interval or one-off. Pick a cadence in plain language and preview the next five runs before you save.

  • Integrations

    Connect Gmail, Google Ads, Google Analytics, Airtable, BigQuery and Notion. Choose which accounts and properties each run can see.

  • Approvals

    Review the actual email, record or change before it goes out. Edit it in place. Approve or decline.

  • Results

    Charts, reports and documents come back as versioned artifacts. Share one by link, without giving anyone an account.

Sixty-six seconds, prompt to approval.

Configure an agent, connect an account, put it on a schedule, watch it work, and approve the one thing that needs a person. No sound.

Built for people who will read the code.

The parts that are easy to get wrong are handled in the platform, not left to each implementation.

  • Code Mode

    An agent can write a short Python script that chains its tools together. It runs in a sandbox with no network or credentials, and every call inside it is still approved and audited.

  • Delegation

    An agent can hand part of a task to another agent you have allowed. The child run inherits the same approval rules.

  • Isolation in the database

    Workspaces are separated with Postgres row-level security, not just application code. Each workspace gets its own storage bucket.

  • Your keys stay yours

    Secrets are referenced from your secret manager, never stored in the app. Praxis sends no telemetry.

See how it works

Three commands to a running workspace.

You need Docker and an API key from OpenAI, Anthropic or Google, or an open-source model served locally. Nothing else has to be installed.

Read the docsView on GitHub
git clone https://github.com/Greg-Asquith/praxis-agents-os.git
cd praxis-agents-os
make quickstart
Licence
Apache 2.0
Telemetry
None
Deployment
Your infrastructure

Want it set up and built around your workflows?

Greg Asquith created Praxis Agents OS and implements it for organisations: deployment, integrations, agents, approval rules, and an ongoing roadmap of private capability. The software stays open source either way.

Work with GregDiscuss an implementation