Get started
Rivet needs Node.js 22 or 24 and Git on macOS or Linux. Your application can use any language or framework.
1. Install Rivet
Install the published alpha package. Choose Yarn Classic, npm or pnpm:
yarn global add @agilno-tech/rivetnpm install --global @agilno-tech/rivetpnpm add --global @agilno-tech/rivetrivet --helpModern Yarn does not provide yarn global; use npm or pnpm for the global CLI instead.
For a pinned version, PATH help or a project-only installation, see installation.
2. Connect your project
From the project root in your regular terminal:
rivet setup
rivet setup --write
rivet doctorThe first command previews the configuration. Review the checks and dependency commands before applying it with --write. Setup installs project policy and Rivet skills for Claude Code and Codex. It preserves existing valid configuration and does not run project checks or install application dependencies.
Rivet proposes checks for supported Node, Django, pytest, Go and Rust layouts. If it cannot find a check, interactive setup asks for your verification command. Use your project's normal tools, including Docker if required. See custom check and dependency configuration for other layouts.
Resolve the required checks shown by doctor. A preparation-required result means a reviewed dependency plan can prepare an isolated checkout later; dependencies are not installed yet. Review and commit setup files, then start from a clean checkout of the configured default branch. Host workflows also require a fresh or ahead remote-tracking ref; rivet preflight --mode=host reports readiness.
If setup cannot write the harness skill, use your regular terminal or approve the exact operation through the harness. See permission troubleshooting.
3. Complete a task
Choose one of these entry points.
In a terminal
Install and authenticate Claude Code or Codex CLI, then run:
rivet run "Add a GET /health endpoint with a regression test"Run from the project root or any folder inside it. Rivet detects the project and offers a choice if both CLIs are available. To select one directly:
rivet run "Add a GET /health endpoint with a regression test" --harness=claude
# Alternatively:
rivet run "Add a GET /health endpoint with a regression test" --harness=codexReview the task, numbered acceptance criteria, proposed files, checks, permissions and limits before approving. Add --details to include full graph identifiers. The activation prompt waits while you read; closing it leaves a saved proposal. To review that proposal again, run rivet task start (or rivet task resume). Rivet works in isolated Git checkouts and asks separately before preparing their dependencies. Existing dependencies in your source checkout are not copied. Installing Rivet does not change your project's package manager; Yarn, pnpm, npm and Bun project scripts remain supported.
In Claude Code, Codex or another coding harness
Open your configured project, reload the harness if needed, and ask:
Read the Rivet skill. Add a GET /health endpoint with a regression test using this project's framework. Show me the plan before starting, work in Rivet's isolated checkout, and show the changed files and executed checks for final review.
The harness manages the detailed Rivet commands and task identifiers. Approve the required file, Git and worktree operations through its normal permission flow. See compatibility for tested CLI paths and current desktop limitations.
4. Review and finish
rivet task statusAfter verification succeeds, rivet run opens final review in the same terminal. Choose:
- Review changes: view the exact diff from the starting commit to the tested commit. Large diffs must be inspected in the reported integration checkout.
- Apply locally: confirm the exact commit. Rivet fast-forwards your original clean default branch to the tested result and records the task as completed. It does not push or deploy. The branch must still be at the task's starting commit, and the committed project policy must be unchanged.
- Pull-request delivery: prepare a local delivery record, then follow the displayed publish and review commands with separate approvals. See delivery for provider setup.
- Leave for later: keep the verified work saved.
To pick up where you left off:
rivet task resumeResume reviews an unapproved plan, continues an eligible interrupted terminal task, or reopens final review. rivet task start and rivet task approve remain available as direct shortcuts. For a task owned by a coding harness, resume shows how to continue through that harness; verified host work can use the same final-review menu.
If your original checkout has changed, preserve your edits and follow the reported guidance; Rivet will not force a merge. Completed tasks can be inspected with rivet task status --run=<id>.
For interrupted terminal tasks or missing dependencies:
rivet task resume
rivet task depsA saved terminal proposal is reviewed before activation; you do not need to call feature start separately. Resume an eligible activated terminal task after correcting the reported problem. For a task owned by your coding harness, continue through that harness. task deps prepares a clean active Worker or accepted integration checkout using its configured dependency plan. Source corrections may require a new reviewed plan. See troubleshooting for blocked tasks.
You normally do not need a task ID or --project. When several tasks are active, an interactive terminal shows a numbered list of task descriptions and states. Choose one, or use --run=<id> explicitly. Noninteractive commands require an explicit selector when the choice is ambiguous. Use --project=<path> when running outside the project.
Report a problem
Human-readable command failures show a next step and the path to a private JSON diagnostic report when it can be saved. Review that file before attaching it to a GitHub issue, along with reproduction steps and expected behavior. To collect broader sanitized diagnostics:
rivet support --saveNothing is uploaded automatically. See report contents and sharing.
Add more when you need it
- Jira, Linear, Figma and Confluence: optional context and ticket connections.
- Project protocols: your team's reviewed procedures.
- Model selection: worker harnesses and advisory text models.
- Runtime reference: configuration and advanced commands.