Topics
Code graphsAI agentsCoding agentsSelf-improving modelsSelf-evolving agentsAutonomous agentsProduct ReleasesLatest
A prepaid wallet for agent spendCloud sandboxes for coding agentsDecisions in one place: the Console and the Work boardEvidence an auditor can checkPaperField manualRSS feedOxagen on Hugging FaceWorks with
Keep the coding agents your team uses. Oxagen brings recorded work from supported agents into one workspace. Capture, cost reporting, live messages, and approvals differ by integration.
The marks below show agents, providers, and integration paths to check for your setup.
Anthropic’s coding agent for the terminal and the IDE.
Oxagen’s hooks capture the prompts, tool activity, and stops Claude Code exposes. Check the support table for live direction and approval coverage.
oxagen agent enroll --harness claude-codeThe GPT models, and the OpenAI Agents SDK for building agents on them.
Oxagen records supported model calls routed through it. For an Agents SDK agent, check the wrapper’s capture coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })LangChain’s library for agents built as graphs of steps that keep state.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })The language most agent frameworks ship in first.
Start a custom agent under oxagen agent run. Oxagen records the activity the integration reports. Controls apply to supported calls routed through Oxagen.
oxagen agent run --name my-agent -- python agent.pyThe AI code editor, and its cursor-agent command line.
Oxagen records the activity Cursor exposes. Model cost capture, live direction, and controls depend on the integration.
oxagen agent enroll --harness cursorVercel’s TypeScript toolkit for building agents and AI apps.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })A Python framework for crews of agents that each take a role.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })The company behind the Claude models.
Bring your own Anthropic key. Oxagen counts and prices each Claude call routed through it, and the run’s budget applies.
For agents built for speed and memory safety.
Start a custom agent under oxagen agent run. Oxagen records the activity the integration reports. Controls apply to supported calls routed through Oxagen.
oxagen agent run --name my-agent -- ./target/release/agentOpenAI’s coding agent for the terminal, the IDE, and the Codex app.
Captured Codex activity joins the workspace’s run records. Check the support table for capture, live direction, and control coverage.
oxagen agent enroll --harness codexAn open-source TypeScript framework for agents and workflows.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })The Model Context Protocol, the open standard agents use to call tools.
Oxagen’s MCP server gives any MCP client Oxagen’s own tools. Enrolling a harness adds it under the name oxagen.
The open-source coding agent Oxagen builds.
Oxagen adds its hook block to Stella’s settings and records each run beside your other agents.
oxagen agent enroll --harness stellaAnthropic’s SDK for building agents on the same loop as Claude Code.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })Typed JavaScript, the language of the AI SDK and Mastra.
Start a custom agent under oxagen agent run. Oxagen records the activity the integration reports. Controls apply to supported calls routed through Oxagen.
oxagen agent run --name my-agent -- node agent.jsThe open-source framework that chains model calls, tools, and retrieval.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })Google’s Agent Development Kit for building and deploying agents.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })For agents that ship as one small binary.
Start a custom agent under oxagen agent run. Oxagen records the activity the integration reports. Controls apply to supported calls routed through Oxagen.
oxagen agent run --name my-agent -- go run ./cmd/agentA framework for agents that answer from your own documents and data.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })Anthropic’s desktop app for Claude.
Claude Desktop reaches Oxagen through the Oxagen MCP server. Each call to Oxagen’s tools is checked and recorded.
A Python agent framework from the Pydantic team, with typed inputs and outputs.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })For agents that live inside the JVM services a company already runs.
Start a custom agent under oxagen agent run. Oxagen records the activity the integration reports. Controls apply to supported calls routed through Oxagen.
oxagen agent run --name my-agent -- java -jar agent.jarMicrosoft’s open-source framework for agents that work together in conversation.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })AWS’s open-source SDK for agents that a model drives step by step.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })Hugging Face’s small library for agents that write and run code.
With a supported wrapper, captured steps join the run record. Rules apply to tool calls routed through Oxagen. Check the support table for capture and control coverage. The wrapper below is illustrative.
oxagen.agent.wrap({ ...your agent code here })Product
Oxagen connects the reason for the work to the recorded runs, instructions, changes, and results behind it.
Work
Follow work from its source to the change. Bring work from supported trackers into one workspace, with its operator, captured runs, instruction versions, pull requests, checks, and cost.
Keep acceptance criteria beside the result. A passing check, a merged pull request, a deployment, and an accepted work item each answer a different question.
Ground the next estimate in work you have already completed. Compare scope, agent effort, and cost, then account for human review and dependencies.
The tracker view uses illustrative data. Confirm connector coverage for your team’s tracker before setup.
Illustrative estimates. Agent minutes exclude human review and waiting time.
Mandates
Turn a useful correction into guidance the team can reuse. Review steering, keep its version history, and see which version a supported run received. Equip supported agents from one shared toolbelt.
Set identity, permissions, tools, steering, and budget in the agent’s mandate. For actions routed through Oxagen, rules allow a request, deny it, or send it to a named person.
Open the Mandates pageRuns
Read the work behind the pull request. Inspect captured prompts, context, tool activity, code changes, and cost with the person who directed the run in view.
Change direction while the work is happening. Send a correction to a supported run and see whether it arrived. Use the record to turn repeated detours into a better brief.
Open the Runs pageKeep the evidence behind the change. A locked record shows later changes, and an export lets a reviewer check its integrity outside Oxagen.
sha256:f2c9…34efAction controls
Make each agent’s permissions a team decision. Its mandate brings identity, tools, steering, rules, and budget together. For actions routed through Oxagen, the record shows which rule answered and who approved the request.
Your agents run in their own environments. For supported actions routed through Oxagen, a rule allows the request, denies it, or sends it to a person. Keep the approval beside the action it authorized.
Agents
release-notesYour systems
Record
| Agent | Request | Answer | Cost |
|---|---|---|---|
| Release agent | Push to main | Denied by no-push-main | $0.00 |
| Bug fixer | Update ENG-219 | Allowed by tracker-writes | $0.0031 |
| Test writer | Run the staging migration | Routed to Priya Shah | $0.00 |
| stella | List invoices | Allowed by billing-read | $0.0012 |
| Hotfix agent | Refund $1,400 | Routed to two people by refund.over-1000 | $0.00 |
| Triage loop | Post to #releases | Allowed by release-notes | $0.0008 |
One view
Keep your active runs within reach. See recorded work from supported agents on one timeline, with the operator, work item, state, and available cost beside each run.
Know who directed the agent and what it produced. Open the run behind a work item, read the captured activity, and send a correction where live direction is supported. Check the delivery status before assuming the agent received it.
Spend
See what agent work produced and what it cost. Follow the bill back to the work, then inspect the runs behind unproductive spend.
Read each cost with its basis: measured, reported, or estimated. Investigate a finding, change the instructions or tools, and compare similar work afterward. Report realized savings separately from the opportunity identified.
Attribution path
$1,848.60 this month across 5 operators and 4 agents
Basis
| From | To | Spend |
|---|---|---|
| Priya Shah | Release agent | $240.10 |
| Priya Shah | Bug fixer | $96.40 |
| Priya Shah | stella | $76.30 |
| Marcus Lee | Bug fixer | $301.70 |
| Marcus Lee | Test writer | $148.20 |
| Marcus Lee | Release agent | $88.30 |
| Ana Ruiz | Test writer | $192.60 |
| Ana Ruiz | stella | $104.50 |
| Ana Ruiz | Bug fixer | $64.30 |
| Jun Park | Release agent | $171.20 |
| Jun Park | Test writer | $83.40 |
| Jun Park | stella | $55.30 |
| Sam Okafor | stella | $128.90 |
| Sam Okafor | Bug fixer | $97.40 |
| Release agent | Merged | $352.40 |
| Release agent | In review | $88.10 |
| Release agent | Returned | $31.60 |
| Release agent | Unproductive | $27.50 |
| Bug fixer | Merged | $391.20 |
| Bug fixer | In review | $71.40 |
| Bug fixer | Returned | $42.90 |
| Bug fixer | Unproductive | $54.30 |
| Test writer | Merged | $268.70 |
| Test writer | In review | $96.80 |
| Test writer | Returned | $18.40 |
| Test writer | Unproductive | $40.30 |
| stella | Merged | $241.60 |
| stella | In review | $62.20 |
| stella | Returned | $24.70 |
| stella | Unproductive | $36.50 |
The costs shown are the costs Oxagen recorded.
Daily spend for 30 days
| Codex | $559.80 |
|---|---|
| Claude Code | $499.60 |
| Cursor | $424.20 |
| stella | $365.00 |
Usage and the rate applied to it
Recorded spend per merged pull request
Roles
Engineers direct the runs, leaders review delivery, and product managers check the result against the work they asked for.
See what agent work produced and what it cost. Follow an item through its captured runs, pull requests, checks, and acceptance. Use the record to find where the work needs to improve.
Give supported agents instructions the team can review and toolbelts the team approves. Keep versions and delivery records together so you can see what reached each run.
| Agent | Mandate | Last run |
|---|---|---|
| Bug fixer | billing-fixes | 4 min ago |
| Release agent | release | 12 min ago |
| Docs writer | docs-only | 1 hr ago |
| Hotfix agent | hotfix | 3 hr ago |
| Test writer | tests | 6 min ago |
Know whose authority is behind the action. For requests routed through Oxagen, inspect the rule that answered and the person who approved it. For mediated connections, the credential stays in Oxagen.
Follow the bill back to the work. Inspect the runs behind unproductive spend, choose a fix, and measure what changed. Keep flagged spend separate from savings you have measured.
Ground the next estimate in work you have already completed. Keep acceptance criteria in view and check the result against the work you asked for. Account for human review and dependencies when planning dates.
Change direction while the work is happening. Steer supported runs from one workspace and check whether your message arrived. Turn a useful correction into guidance the team can review and reuse.
| Run | State | Questions |
|---|---|---|
| Invoice preview shows usage lines | Waiting on you | 2 |
| Retry the Stripe usage sync | Working | 0 |
| Rate-limit meter writes | Done | 0 |
Industries
Keep the source work, recorded changes, and applicable approvals together for the reviews your team needs to make.
Several agent tools across many repositories, each with its own history, and a backlog longer than the team.
A review of the billing service needs the recorded agent work behind the code changes.
Code near patient data needs a clear account of who requested a change and which controls applied.
Carriers that price agent risk ask for evidence of authority and governed actions for a defined workload.
Research
Read the research behind longer agent runs, live direction, and the human review that turns a code change into accepted work.
An agent can write a change in minutes. Then the change waits for a person to read it. This post measures that wait.
When a long run goes wrong, most teams can stop it or type at it. A steer is a message with a delivery mode, a status, and a record.
METR measures how long a task an agent can finish on its own. That length has doubled about every seven months since 2019.
Field manual
The field manual runs to 21 parts. Leave your email address to read it.
Pricing
Start with one team and follow its agent spend through to the work it accepts.
Team and Business seats include monthly agent actions and credits. Usage beyond the allowance draws on prepaid credits. Your model provider bills calls made on your own key.
Prices in USD. Checked 3 November 2026.
Compare the plans| Run | Agent | Cost |
|---|---|---|
| Fix invoice preview | Claude Code | $2.57 |
| Review usage lines | Codex | $2.49 |
| Address review | Claude Code | $1.41 |
Get a demo
Choose a repository and a piece of work to discuss, then trace the work with the team in a supported setup.