# The control plane for AI engineering | Oxagen

Source: https://oxagen.sh/product

The whole product on one page

# Product overview

Oxagen connects coding-agent runs to the people, work items, code changes, costs, and outcomes behind them. Engineering teams use that record to direct supported agents, share reviewed steering and toolbelts, enforce controls on mediated actions, and improve how the next task gets done.

[Follow work from its source to its outcome](https://oxagen.sh/#demo)

The Runs page with illustrative data. Twenty runs are live across Claude Code, Codex, stella, and Cursor, two wait on a person, and the table lists each run with its state and its cost so far.

The Runs page with a workspace's live runs

01

## On this page

*   [Work](https://oxagen.sh/product#work)
*   [Mandates](https://oxagen.sh/product#mandates)
*   [Runs](https://oxagen.sh/product#runs)
*   [Setup](https://oxagen.sh/product#setup)
*   [Supported agents](https://oxagen.sh/product#supported-agents)

02

## Work

A reviewer opens a pull request and asks why it exists. The answer sits in a tracker, in a run on someone's laptop, and in a checklist nobody wrote down. Oxagen puts them on one work item.

Triage suggests a priority and a first Done Checklist, the short list of things a reviewer checks. A person approves the checklist and sends the item to one agent. When the run ends, the item shows the pull request, its head commit, and the checks the base branch requires on that commit.

Accept records a person's acceptance of that head commit and merges nothing. The item is done once a person has accepted it and the pull request has merged, in either order. A pull request closed without merging is not done.

The work item WI-12, Retry the export when the API answers 429, with illustrative data. Its Done Checklist has 2 of 4 items done, two runs worked on it, and its pull request acme/platform#612 is open with passing checks.

A work item with its Done Checklist and its pull request

03

## Mandates

For a call routed through Oxagen, a rule your team wrote allows it, denies it, or routes it to a named person. For a mediated connection, Oxagen uses the connection credential on the agent's behalf, and the agent does not receive it. In observe mode, Oxagen records what the policy would decide and lets the call go ahead.

An agent's identity, permissions, toolbelt, steering, and budget together make up its mandate. Oxagen checks those terms on calls routed through it and records each decision, so the people who set them can see how they applied.

The rules table on the Guardrails page, with illustrative data. Small refunds to known customers approves Stripe refunds of $500.00 or less and has approved 212 calls and held 9. Weekday dependency bumps has approved 64 and held 2, and Repeat deploys to staging is off.

Rules on the Guardrails page

04

## Runs

An agent that works for an hour leaves a scrollback nobody reads. Oxagen records the run while the agent works: each prompt, each tool call and the rule that answered it, the files git reports as changed, and the cost of each model call it carried.

Know whose authority is behind the action. The record names the agent that asked, the person it acted for, the rule that answered, and who approved the call when a person was required, beside its result and cost.

The record holds only what the agent's computer reported, and the [coverage statement](https://docs.oxagen.sh/docs/security/record-coverage) names what each harness reports. Cursor's model calls do not pass through Oxagen, so their cost is not recorded.

A live run's page with illustrative data, printed the way stella prints it. The run, Credit a paid invoice once per webhook event, shows its prompt and each tool call in the transcript, with its tasks, changes, and context beside it.

A live run printed the way stella prints it

05

## Setup

1.  **Install the CLI**
    
    Install the Oxagen app on the machine the agent runs on. It puts the oxagen command on your path.
    
2.  **Enroll the agent**
    
    Run `oxagen agent enroll --harness claude-code` or name Codex, Cursor, or Stella. Enrollment writes the harness's hooks and gives the machine its own identity in the workspace.
    
3.  **Confirm the first run**
    
    Start the agent and open Runs. The run appears while the agent works.
    

06

## Supported agents

Controls Oxagen applies to each supported agent
| Control | Claude Code | Codex | Cursor | Stella |
| --- | --- | --- | --- | --- |
| Record each tool call | Yes | Yes | Yes | Yes |
| Let a rule allow or deny a tool call | Yes | Yes | Yes | Yes |
| Send a tool call to a person | PartialClaude Code asks you on your own machine. | NoOxagen denies the call and says why. | NoOxagen denies the call and says why. | PartialStella holds the call for your approval. |
| Record model calls and their cost | Yes | Yes | NoCursor sends model calls to its own servers. | PartialFor Anthropic models. |
| Send a message to a running agent | Yes | Yes | PartialDelivered at the end of the turn. | Yes |
| Stop a run | Yes | PartialOxagen blocks the next tool call and ends the model call in progress. | PartialOxagen blocks the next tool call. | Yes |

07

## Next steps

*   [Mandates](https://oxagen.sh/product/mandates)
*   [Runs](https://oxagen.sh/product/runs)

## Related pages

*   [
    
    ### Mandates
    
    The terms an agent works under
    
    ](https://oxagen.sh/product/mandates)
*   [
    
    ### Runs
    
    The record each agent run leaves
    
    ](https://oxagen.sh/product/runs)

## Product walkthrough

Bring one piece of work and see how its runs, controls, and record fit together.

[Follow work from its source to its outcome](https://oxagen.sh/#demo)
