Create or sign in
Use the web console to create a workspace. Your workspace is where hosts, sessions, approvals, and audit history are grouped.
Documentation
Connect agents, route approvals, and keep a clean audit trail.
UpSteer gives local and cloud agent runs a shared control plane. Use these docs to install the CLI, link a host, start sessions, and decide which actions should pause for human review.
terminalpublic setup
curl -fsSL https://upsteer.ai/install.sh | sh
upsteer start codexSession events, approvals, and runtime health stream to your workspace.
Quickstart
Run your first controlled agent session
Install the CLI
Install UpSteer on the machine where your agent runs. The CLI opens a browser sign-in flow or shows a short code for SSH environments.
Start an agent
Run an agent through UpSteer. The session appears in the web and mobile console with live progress and approval cards.
Core concepts
The model is intentionally small
Workspace
The account boundary for hosts, live sessions, rules, billing, and audit history.
Host
A linked machine or runtime that can start agent sessions and report health.
Session
A single agent run with status messages, tool activity, decisions, and completion state.
Approval
A sensitive action that pauses for allow, deny, or rule evaluation before continuing.
Commands
Common commands
curl -fsSL https://upsteer.ai/install.sh | shInstall the UpSteer CLI on a workstation or devbox.
upsteer start codexStart a Codex session with live status and approval routing.
upsteer start claudeStart a Claude Code session connected to your workspace.
upsteer start geminiStart a Gemini CLI session with the same control layer.
upsteer start opencodeStart an OpenCode session and stream activity into UpSteer.
Approvals
Pause only when judgment matters
UpSteer is designed for the moments where a human should stay in the loop: deploys, file writes, dependency changes, external network calls, credential-adjacent operations, and other actions with real blast radius.
Each approval captures enough context to decide quickly: the agent, host, command or action, working directory, risk, and recorded outcome.
Recommended default
Keep routine read-only activity flowing, and require approval for writes, deploys, destructive commands, and production-impacting operations.
Rules
Automate the boring decisions
Allow trusted patterns
Let predictable commands continue without interrupting you, such as local test runs or read-only inspection.
Keep risky actions gated
Require manual review for production deploys, migrations, package changes, and commands that can delete or overwrite data.
Review the audit trail
Use history to tune rules over time and understand which decisions were made manually or automatically.
Troubleshooting
When a session does not appear
The host is not linked
Run the start command again and complete the browser sign-in or enter the short code in the web console.
The agent is running on SSH
Use the code shown in the terminal. This avoids needing a browser on the remote machine.
Approvals are not arriving
Check that the host is online, the session is still active, and your approval rules are not automatically resolving the action.
Visibility
What stays public vs. private
These docs are public because install, setup, and product education help adoption. Account-specific tokens, workspace IDs, private audit data, internal runbooks, and any proprietary implementation details belong behind authentication.