Review, merge, deploy, monitor, respond to incidents, track issues — plain CLI commands, scoped to exactly the actions you allow.
Works with every coding agent — if it can run a command, it can use OmniCLI.
Build a definition in the dashboard — anything you leave unchecked simply doesn't exist to your agent.
Anything you didn't select simply doesn't exist in your CLI — the agent gets unknown command.
Every definition is cryptographically signed, so the surface your agent sees is exactly the one you approved — tamper-proof and enforced client-side.
--help is for humans. Your agent reads omni-cli discovery — JSON with the exact parameters and output schema for the action it's about to run.
the index — only the platforms your definition allows
the full schema — fetched only for the action it will run
One install works with every platform — instead of a separate server, config, and token for each tool.
Your signed definition decides. Unselected actions don't exist in the CLI — not disabled, absent.
Signing in, signing out, and account changes are yours alone — your agent can't run them.
Your agent looks up an action right before using it, instead of carrying every tool's description the whole time.
Every platform is already modelled. No servers to host, no schemas to write.
One dashboard lists every run across your platforms — what ran, whether it worked, and which actions changed something.
One install per person. Update the definition; connected machines follow.
MCP and OmniCLI aren't exclusive — run a server where you need a long-lived, stateful connection to one system, and one command line for everything else.
The browser handles login — never the agent. Here's the full GitHub flow, verbatim.
Device flow: visit the link, enter the code. The browser handles login, not the agent.
Authorize GitHub once — your token is validated and stored locally.
Ask in plain language — the agent picks the scoped command and runs it. No tokens pasted, no secrets in prompts.
OmniCLI is the boundary between your agent and your tools — and it's built to keep secrets out of prompts.
Every token stays on your machine — never sent to agents, pasted into prompts, or off your device.
Agents run only the platforms and actions you turned on — like omni-cli github pr diff — and nothing more.
Each action is a signed, auditable CLI command with a fixed surface — no raw API access or arbitrary calls.
See the audit trail →A failed action runs once and returns structured context, so your agent decides the next step — no duplicate PRs or deploys.
Every run is recorded — which platform, which action, and whether it succeeded — so giving your agent real capabilities never means losing sight of how they're used.
pr merge and deployment deploy are flagged, so side effects are easy to spot.Write one command and it runs identically whether it's Cursor, Codex, or Claude Code calling it. No per-agent setup, no glue scripts.
pr create and issue transition mean the same thing everywhere.Sign up, pick your platforms and actions, and get a copy-paste install one-liner for your agents.