# Oxagen 5.5.0, the last beta · Oxagen Research

Source: https://oxagen.sh/blog/oxagen-5-5-0

[Research](https://oxagen.sh/blog) · [Product Releases](https://oxagen.sh/blog/pillars/product-releases)

# Oxagen 5.5.0, the last beta

Oxagen 5.5.0 is the last beta, a pre-production release. It runs coding agents in cloud sandboxes, prices model calls before they run, and puts every decision in one place.

Oxagen Research · November 5, 2026 · 6 min read

*   [Product Releases](https://oxagen.sh/blog/pillars/product-releases)

A team that runs coding agents every day knows the morning routine. One agent stopped overnight when a usage limit hit. Another asked a question in a thread nobody was watching. A CI failure sat in an inbox for an hour. Somebody has to open six tabs to learn what happened and what it cost.

Oxagen 5.5.0 is a beta pre-production release. It is the last beta before general availability. The release moves agent runs off the laptop, prices each model call before it runs, and gathers the decisions a team owes its agents into one queue. The full notes are on the docs site.[1](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-1)

## Cloud sandboxes

A coding agent on a laptop stops when the lid closes or the plan's usage limit hits. In 5.5.0, a person starts a Claude Code, Codex, or Stella run in a sandbox on the organization's own Modal account.[2](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-2) The run starts from the app, a stella thread, `oxagen sandbox run`, or the `launch_sandbox_run` MCP tool. A send from a stella thread waits as an approve or deny card. Each run checks out every repository the workspace links. A Claude Code run that hits every saved account's usage limit waits for the reset and continues the same session. A headless Stella run ends instead. A run stops at its work item's time limit, and the terminal drawer pauses the agent before anyone types.[3](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-3) Several agents share one sandbox only by taking turns, and only when the workspace turns that on.

We shipped this because a run that sees one repository, or dies at a five-hour limit, leaves work half done. The details and limits are in [Cloud sandboxes for coding agents](https://oxagen.sh/blog/cloud-sandboxes-for-coding-agents).

## Prepaid wallet

Before this release, Oxagen priced a model call after it ran, so a call could spend credits the organization did not have. Now each model call on Oxagen's key costs a fixed 1, 15, or 30 credits by size, charged before the call runs. The month's Team or Business allowance pays first, then the prepaid wallet. An empty wallet or a failed seat payment holds new runs, and started runs finish under their rules. Billing owners get an in-app notice and an email at 50, 80, and 100 percent of the month's allowance.[4](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-4) Credit refunds, a billing portal session, and sandbox slot purchases are available through the API and MCP.

Size pricing ignores the token count, and alerts come from a daily sweep. [A prepaid wallet for agent spend](https://oxagen.sh/blog/a-prepaid-wallet-for-agent-spend) covers both.

## Console and Work board

Agents raise questions at any hour, and results pile up for review. The Console now asks each question as a set of choices, with stella's recommended answer marked and nothing picked for you.[5](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-5) While you were out lists what stella did on its own, each with an undo button. CI results, merged pull requests, failed runs, and refused model accounts arrive as cards you read once.

Work items also show as a board. Dropping a card opens the action the move means, such as approve, send, retry, or close, and the card moves once the write lands.[6](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-6) Reviewers accept results in bulk. stella, MCP, and the API add, edit, delete, or re-triage up to 100 items per call, and the CLI reaches the same tools through `oxagen mcp call`.[7](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-7) One limit stands out: an enrolled machine can create and edit items without an approval card.

[Decisions in one place: the Console and the Work board](https://oxagen.sh/blog/decisions-in-one-place-the-console-and-the-work-board) explains why each card works the way it does.

## Run page

A long run can bury one failing test under hundreds of steps. The new Run page has two columns, a sticky header, and the agent's task plan under Run details. The transcript reads as a conversation, or as a timeline that replays from the first step. A Problems panel lists loops, groups of refusals, setup failures, and failing MCP tools, each linked to its first call. A refused model call shows its model and reason. When a wrapped agent starts to loop, it hears about it once.[8](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-8) The notice does not pause the run.

[Reading a long agent run](https://oxagen.sh/blog/reading-a-long-agent-run) covers the loop window and what the record depends on.

## Model credentials

A person who worked in two organizations used to save each key twice. A model credential now belongs to the person who saved it and serves their runs in every organization they work in. A workspace account lets an agent's sandbox runs bill the workspace. On an enrolled machine, a run moves to the next saved credential when one hits a usage limit. A refused login is signed in again on the owner's own machine with `oxagen credentials reauth`, and `oxagen workspace use` points a device at one workspace.[9](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-9) Cursor is not covered yet.

The MCP Studio Test tab now opens with sample arguments in each tool's input shape, so a first test call needs no typing. [Model credentials that follow the person](https://oxagen.sh/blog/model-credentials-that-follow-the-person) states the limits on copies and shared logins.

## Steering

An agent refused the same command run after run used to leave no rule behind. stella now drafts a steering record only from what a person corrects after the first turn, and a model judge passes each draft. A nightly pass turns repeated refusals, recurring setup failures, and failing tools into fix cards on the Console. A person can add a skill from skills.sh through a steering pull request.

Steering reads now answer only the app, stella, and runs of Claude Code, Codex, Cursor, and Stella that Oxagen wraps through each harness's hooks.[10](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-10) A key alone no longer reads a workspace's steering. That is a breaking change for API and MCP callers that read steering with a key. [Steering drafted from corrections](https://oxagen.sh/blog/steering-drafted-from-corrections) lists who is affected.

## Evidence

After an agent incident, an auditor or an insurer asks what the record proves. `oxagen verify` checks a run export offline, so the auditor does not have to trust the vendor that kept it.[11](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-11) The claim bundle format adds RFC 3161 timestamps from a third party.[12](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-12) A control-by-control mapping to AIUC-1 marks 26 of its 140 controls covered by Oxagen's record and 33 partly covered.[13](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-13) It leaves 5 missing and 76 for the customer to answer. Managers can rank operators from the record and read one person's scorecard.

The product does not yet export a claim bundle for a real agent run, only a sample. A passing check does not prove that every action was recorded. [Evidence an auditor can check](https://oxagen.sh/blog/evidence-an-auditor-can-check) explains both.

## Dashboards and Spend

Reports now lists 31 dashboards, each with the questions it answers. They cover spend and outcomes, unproductive spend, task completion, model and effort fit, ready work, CI, and the codebase. The Console shows a few by default. Spend moved under Reports and takes any set of workspaces, a calendar range, and a comparison with an earlier range. It reads one stored row per root run, so a page view no longer loads every run in range. The codebase dashboards need the code graph. The operator ranking and prompt habits dashboards open only to Owners, Compliance and Billing, and the Admins of the workspace they read.

## Design system

The app, the docs site, and the component gallery now draw from one shared component kit and one stylesheet, so what a designer reviews is what ships. Pages follow a quiet layout: each one answers one question at the top and keeps the rest behind a menu, a fold, or a drawer. The type moved to Geist, and the palette moved to cool neutrals with one gold.

## Hosting

The app, the API, and the Postgres database run in Google Cloud's us-east4 region in Virginia.[14](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fn-14) Usage telemetry and background jobs run on hosted services outside Google Cloud. Every commit to the main branch that passes its checks deploys to production.

## References

1.  Oxagen (2026). _Releases_. Oxagen documentation. [https://docs.oxagen.sh/docs/releases](https://docs.oxagen.sh/docs/releases) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-1)
    
2.  Modal (2026). _Sandboxes_. Modal Docs. [https://modal.com/docs/guide/sandbox](https://modal.com/docs/guide/sandbox) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-2)
    
3.  Oxagen (2026). _Agent sandboxes on Modal_. Oxagen documentation. [https://docs.oxagen.sh/docs/agent/sandboxes](https://docs.oxagen.sh/docs/agent/sandboxes) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-3)
    
4.  Oxagen (2026). _Billing and usage_. Oxagen documentation. [https://docs.oxagen.sh/docs/billing](https://docs.oxagen.sh/docs/billing) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-4)
    
5.  Oxagen (2026). _The Stella page_. Oxagen documentation. [https://docs.oxagen.sh/docs/stella-page](https://docs.oxagen.sh/docs/stella-page) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-5)
    
6.  Oxagen (2026). _Work_. Oxagen documentation. [https://docs.oxagen.sh/docs/work](https://docs.oxagen.sh/docs/work) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-6)
    
7.  Oxagen (2026). _MCP tools_. Oxagen documentation. [https://docs.oxagen.sh/docs/mcp/tools](https://docs.oxagen.sh/docs/mcp/tools) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-7)
    
8.  Oxagen (2026). _Wrap an agent_. Oxagen documentation. [https://docs.oxagen.sh/docs/cli/wrap-an-agent](https://docs.oxagen.sh/docs/cli/wrap-an-agent) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-8)
    
9.  Oxagen (2026). _CLI commands_. Oxagen documentation. [https://docs.oxagen.sh/docs/cli/commands](https://docs.oxagen.sh/docs/cli/commands) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-9)
    
10.  Anthropic (2025). _Hooks reference_. Claude Code Docs. [https://code.claude.com/docs/en/hooks](https://code.claude.com/docs/en/hooks) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-10)
     
11.  Oxagen (2026). _Export and verify a run_. Oxagen documentation. [https://docs.oxagen.sh/docs/security/export-and-verify](https://docs.oxagen.sh/docs/security/export-and-verify) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-11)
     
12.  Adams, C., Cain, P., Pinkas, D., & Zuccherato, R. (2001). _Internet X.509 Public Key Infrastructure Time-Stamp Protocol (TSP)_. IETF RFC 3161. [https://www.rfc-editor.org/rfc/rfc3161](https://www.rfc-editor.org/rfc/rfc3161) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-12)
     
13.  AIUC (2026). _AIUC-1_. Artificial Intelligence Underwriting Company. [https://standard.aiuc-1.com](https://standard.aiuc-1.com/) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-13)
     
14.  Google Cloud (2026). _Regions and zones_. Compute Engine documentation. [https://cloud.google.com/compute/docs/regions-zones](https://cloud.google.com/compute/docs/regions-zones) [↩](https://oxagen.sh/blog/oxagen-5-5-0#user-content-fnref-14)