Turn what you know
into what gets done.
Praxis Agents OS is where your team's knowledge, data and ways of working live, on servers you control. Agents put them to work across the systems you already use, and check with a person before anything important changes.
Northwind delay reply
Gmail Agent · orders@yourcompany.co.uk
A working illustration built from the product's real run states and tool names. Nothing you approve here is sent anywhere.
Six runs, six industries, all ten integrations. Each agent starts from what the business already knows, does the work, checks the one change that matters with a person, and shows the finished result. Real tool names, fictional companies.
What you know stays yours.
Your know-how should not live inside someone else's product. In Praxis it sits in your database, on your servers, under an open licence. Change the model and keep the knowledge, the instructions and the history.
Knowledge
Documents, pages and files in a knowledge base that agents search and cite. Notion pages can feed it too.
search_knowledgeKnow-how
Skills: how your team does a job, written once with its reference documents, and used by every agent that needs it.
read_skill_documentMemory
Agents keep notes as they work, each with where it came from. You review, correct, archive or purge them.
save_memoryWays of working
Agents, instructions, schedules and context groups. Your repeat work, written down as settings you can read.
schedulesData and results
In your Postgres and your storage bucket. Reports and charts come back as versioned artifacts that stay there.
create_artifact
Put to work where your data already lives.
Ten systems your team already uses, and any model you like. Connect an account once, then decide which ad accounts, mailboxes, drives, bases and datasets each run is allowed to see. Outlook, SharePoint, Search Console and Meta Ads are the newest.
Gmail3 tools
Outlook Mail12 tools
SharePoint9 tools
Google Ads22 tools
Meta Ads1 tools
Google Analytics5 tools
Search Console5 tools
BigQuery3 tools
Airtable4 tools
Notion6 tools
OpenAI
Anthropic
Google Gemini
Azure OpenAI
Open-source models
Say what you want.
Watch every step.
No black box. The sources it read, the plan, the tools, the accounts and the results are on screen while the agent works, and afterwards.
- 01
You ask
In plain English, or on a schedule. “Every Monday at 07:00” is a valid instruction.
- 02
It reads what you know
It searches your knowledge base, follows your skills, checks its notes, and cites the sources it used.
- 03
It does the work
Reports, queries, drafts and records, across the accounts you gave it. Every tool call is on screen.
- 04
It checks with you
Where your rules say so, it stops and shows you the real email, record or change before it goes anywhere.
- 05
It keeps the result
Reports, notes and an audit row go back into your workspace, so the next run starts from more.
The Monday report, done by 07:00.
Put the work that comes round every week on a schedule. Agents do it the way your team does it, from what your team wrote down, while you get on with something else.
- Mon 07:00The Monday reportReporting AgentReport ready to review
- Mon 08:30Inbox draftsGmail Agent3 sent after approval
- Tue 08:30Inbox draftsGmail Agent2 sent, 1 declined
- Tue 10:00Search term clean-upGoogle Ads Agent14 negatives proposed
- Wed 08:30Inbox draftsGmail Agent2 drafts waiting
- Wed 14:20“What did we spend on Brand in Q3?”BigQuery AgentChart returned, SQL shown
- Thu 08:30Inbox draftsGmail AgentReading 6 threads
- Fri 08:30Inbox draftsGmail AgentQueued
- Fri 16:00Week-to-date pacingReporting AgentNext run in 2 days
The Monday report
Pulls Google Ads, Meta Ads and Analytics, writes the commentary the way your team writes it, and leaves it for you with your coffee.
Search term clean-up
Finds the wasted spend against your client notes, proposes the negatives, and changes nothing until someone says yes.
Questions for the warehouse
Ask BigQuery a question in English. Get a read-only query, the SQL it ran, and a chart.
The inbox
Reads the thread, checks your policy, drafts the reply, and holds it until you press send.
Your rules decide what it does alone.
Putting your know-how to work only helps if you trust what the agent does with it. Every tool has a switch: Off, Approval or Auto. Reads can run on their own. Writes wait for a person. Some, like any change to a Google Ads account, cannot be set to Auto by anyone, and the rule lives in code.
- 70
- tools across ten integrations
- 31
- writes that can never be set to Auto
- 400
- days of audit history per call
Tool access
Operations Agent · 8 of 70 tools shown
Try setting a locked write to Auto.
google_ads_run_reportReads
Google Ads
google_ads_add_campaign_negative_keywordsWrites
Google Ads · locked to Approval
google_ads_create_campaign_budgetWrites
Google Ads · locked to Approval
sharepoint_write_fileWrites
SharePoint · locked to Approval
gmail_send_messageWrites
Gmail
notion_create_pageWrites
Notion · locked to Approval
airtable_update_recordWrites
Airtable
bigquery_run_queryReads
BigQuery
One workspace for what you know and what gets done.
What you know, the agents that use it, their schedules and their connections. Results and anything waiting for you turn up where the work happened.
Knowledge base, skills, files and memories. What your business knows, in one place. Agents search and cite it, and you can correct what they remember.
Real screens from a demo workspace. Click to see full size.
Built for the person who reads the source.
The parts that are easy to get wrong are handled once, in the platform, instead of being left to every implementation.
Code Mode
An agent can write a short Python script that chains its tools. The script runs sandboxed, and every call inside it still follows your rules and is audited.
no network · no credentials · 60 s · 25 callsDelegation
An agent can hand part of a job to another agent you allowed. The child run inherits the parent's rules.
delegate_to_agentIsolation in the database
Workspaces are separated by Postgres row-level security, not by trusting application code. Each gets its own storage bucket.
role praxis_app · FORCE ROW LEVEL SECURITYYour keys, your servers
Secrets are referenced from your secret manager, never stored in the app. Nothing phones home. Tracing is off until you turn it on.
AGENT_TRACING_ENABLED=false
Three commands. Your servers. Your keys.
You need Docker and a key from OpenAI, Anthropic or Google, or an open-source model served locally. Nothing else to install, and no account with us.
Receipts for everything. Kept by you.
Every tool call goes into an append-only audit log in your database: the agent, the person who asked, the tool, the provider, the request and the outcome. Over time it becomes the record of how your business actually gets things done.
Here is the same idea applied to your visit. Run the draft at the top, or try forcing a locked tool to Auto, and watch it land.
Audit log · this visit
0 rows · 0 decisions · 0 refused
These rows live in this browser tab only. They are not stored or sent anywhere.
Want it built around your business?
Greg Asquith wrote Praxis Agents OS. He also sets it up for companies: deployment, integrations, getting what your team knows into it, the agents that use it, and the private capability you need next. The code stays open source either way.