Image: GitHub
Why it mattersA team stops choosing one coding agent and can route the same task to whichever one fits, with a single session record to audit later instead of scattered chat histories.
useAgent is an open-source platform that gives Claude Code, Codex and OpenCode a shared place to run: each run happens inside an isolated Linux sandbox with its own terminal, browser and recorded desktop, and every event goes into a Postgres log the team can replay. The repository has 277 GitHub stars, gained in the seven days our feed has tracked it.
One control plane, three engines
The three engines the project supports today are Claude Code, Codex (through either an API key or a ChatGPT subscription) and OpenCode. The README describes the design as "any agent, one interface", with a provider-neutral event contract so a run can be handed to whichever engine fits without rewriting the prompt or the tool set. Runs are durable and event-sourced in Postgres 16 with the pgvector extension, so a restart of the control plane does not lose the session.
Sandboxes come from one of two providers. Daytona is the managed option that pairs with a host on any cloud. CubeSandbox is the self-hosted runtime for teams that want the code and data to stay on their own hardware. Docker is used for the control plane itself rather than as a sandbox choice.
Where the credentials are kept
Credentials never enter the sandbox. The README states this plainly: integrations reach the outside world through what the project calls a "trusted gateway", exposed to the agent as a typed tool while the keys stay on the control plane. Native integrations are listed for Slack and GitHub App auth. Gmail, Linear, Notion and HubSpot arrive through the connector layer.
Native artifacts are handled in place: DOCX, XLSX, PPTX and PDF have their own renderers with revisioned editing, so the agent can hand back a spreadsheet or a slide deck without going through a lossy export.
What it costs to adopt, and what the label says
The project is AGPL-3.0-only, with a separate commercial licence available for teams that need to embed it in a proprietary product or sell it under their own brand. The stack runs on Bun and TypeScript. Reference deployment covers AWS, Google Cloud, Azure, Hetzner and bare metal, with a Terraform script for Hetzner and a provider-agnostic install script for everything else.
The label on the top of the README reads "Alpha software". The authors add that "APIs and schemas may change between releases", so anything built against the current version needs to be prepared for that. The 1,380 commits on the main branch show the pace of change is fast rather than settled.
The interesting idea is that useAgent stops the choice of coding agent from being a per-team standard. A run that suits Claude Code goes to Claude Code, a run that fits Codex goes to Codex, and the record of what happened is kept in one place instead of three separate chat archives. The audit trail is the main benefit: a durable event log makes it possible to answer "what did that agent actually change" a week later without asking the person who ran it to remember.
The trade-off is the amount of infrastructure to run. A control plane with Postgres, a sandbox runtime, a trusted gateway and integrations to Slack, GitHub, Gmail, Linear, Notion and HubSpot is more components than a desktop app talking to a model API, and the alpha label means the schema of that log is still changing. Teams weighing it against a hosted single-agent product should count that cost before the credential-isolation and multi-engine benefits.
Source
Primary source: useagenthq/useagent on GitHub.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
