From a new agent to an audited action, in seven screens
This is the product as it runs today. Every screenshot below is taken from a Praxis Agents OS workspace.
Public build, first tagged release pending. The platform runs end to end with the Docker quickstart. Public cloud deployment guides, a production email transport, self-service password reset, and user-developed applications are still to come.
- 01
Create an agent and give it a brain
An agent is a name, a purpose, written instructions, a model and a set of tools. Pick OpenAI, Anthropic, Google or an open-source model, and a tier from light to max. The exact model sits under Advanced, so most people never need to think about it.
Agents can also carry skills (reusable instructions with reference documents) and be allowed to delegate to other agents.
- 02
Decide what it may do, and what needs a person
Every tool the agent can use has a switch: Off, Approval, or Auto. Reads such as running a report default to Auto. Writes default to Approval. Some writes, such as any change to a Google Ads account, cannot be set to Auto by anyone.
The policy lives with the tool definition in code, so an agent cannot talk its way past it.
- 03
Connect your accounts and choose what each run can see
Connect Gmail, Google Ads, Google Analytics, Airtable, BigQuery and Notion. Praxis discovers the accounts, properties, bases and datasets inside each connection. For every conversation or schedule, you pick which of those the agent works with.
Group resources into named context groups, such as all of one client's ad accounts, and reuse them.
- 04
Give it your knowledge, and let it keep notes
Add documents, pages and files to a knowledge base that agents search and cite. Agents also keep memories, with a record of where each one came from. You can review, correct, archive or purge them.
Retrieval combines keyword and semantic search. Agents cannot write to the knowledge base themselves.
- 05
Run it on a schedule
Recurring, interval or one-off, in the timezone you choose. Pick a cadence in plain language, preview the next five runs, and attach a prompt and the context the run should use.
Scheduled runs pause for approval on external writes by default. Nothing goes out unattended unless the schedule is explicitly allowed to.
- 06
Approve the action, not a blob of JSON
When a run reaches a tool that needs approval, it stops and shows you the real thing: the email with its recipients and body, the record about to change, the campaign about to be paused. Edit it if you want. Then approve or decline.
The run resumes where it left off, even after a restart.
- 07
Keep what came back, and who did what
Charts, reports and documents come back as artifacts with numbered versions. Every tool call is written to an append-only audit log with the actor, the person who asked, the tool, the provider, the request and the outcome.
Audit rows survive the deletion of the agent, the user or the workspace they describe.
The layer around the model, not a replacement for it
Praxis does not compete with model providers or agent frameworks. It gives the agents they produce somewhere to live: identity, permissions, connections, schedules and a record.
- Model providersOpenAI, Anthropic, Google or open-source models supply the reasoning. Swap them per agent.
- Agent runtimePydantic AI runs the model loop: instructions, history, tool calls, streaming.
- Praxis Agents OSWorkspaces, roles, tools and their policies, approvals, schedules, files, knowledge, memory, artifacts, audit.
- Your implementationThe agents, instructions, connections, rules and workflows specific to your organisation.
You run it, so you own the operations
The Docker quickstart gets a workspace running in a few minutes. A production deployment on your cloud needs object storage, a secret manager and someone responsible for keeping it healthy.
Read the setup guide ->- 01Infrastructure: Postgres, object storage, a secret manager, and the API, worker and web containers
- 02Provider accounts and API keys for the models and integrations you enable
- 03Updates, backups, monitoring and incident response
- 04Roles, tool policies and approval rules for your workspaces
The architecture notes explain the why
Nine documents cover the runtime, streaming and durability, governance, context, Code Mode, integrations and the threat model. They describe what is built, not what is planned.