Tiden is live on Product Hunt. See the launch.
Tiden CLI
The tiden CLI checks AI-written code against what you asked for: your agent starts each session from a working contract and can't close it over unverified work. One command wires Claude Code, Codex and your repository, and merging stays with you.
The session loop
The loop commands open and end with the verdict and name the next move, in JSON when piped. The merge stays your call.
The start-intent skill runs tiden intent start with the task you gave it. The working contract lists the specs in scope, how many more the change may affect, and the ones no test verifies. The session's promises are frozen right then, so the close is measured against what was asked before any code existed.
tiden run exec -- go test ./... runs your Go suite unchanged, reports every result and stamps the run with the open session. Any tool that writes JUnit XML reports through tiden run report, and Tiden has reporters for Playwright, Vitest, Jest, PHPUnit and Swift. tiden intent progress shows the agent its verdict, spec by spec.
tiden intent close computes the session's verdict and refuses unverified work with exit code 3. The only way past a red verdict is on the record: a risk acceptance that names the spec and the reason, or a proposed test. The close prints a review card with the branch and its diff, and the merge stays your call.
tiden gate check stops a deploy step
Getting started
Paste the install line, or use Homebrew or PowerShell. It checks the SHA256, and later releases install themselves in the background.
Approve the CLI in your browser. Setup binds the repository to its product, then wires the agent hooks and the 6 skills.
Give your agent a task. It runs tiden intent start and builds against the working contract it gets back.
The close prints a review card with the branch and its diff. You merge in Tiden, or ask your agent to.
Claude Code and Codex get the 6 skills (start-intent, product-context, capture-intent, generate-tests, link-coverage, triage-issues), a hook that asks the agent to open a session, and a Stop hook that sends back a claim of done while the session's ledger is incomplete. Cursor, Gemini and Copilot sessions are captured through hooks and git. Any MCP-capable agent reaches the same specs and gates through the MCP server.
After you approve the CLI in the browser, it stores your login in ~/.tiden/config.json, binds the repository to a Tiden product, wires the agent hooks and a git post-commit hook, and installs the skills. It asks before it adds Claude Code permission rules or Codex network access. tiden uninstall --dry-run lists every file the uninstall would delete or edit.
They stay as they are, in your repository, run by their own runner. Go suites report through tiden run exec -- go test ./..., any tool that writes JUnit XML reports through tiden run report, and Tiden has reporters for Playwright, Vitest, Jest, PHPUnit and Swift. The link-coverage skill then links each test to the specs it verifies.
Only on the record. tiden intent close refuses unverified work and exits 3. The way past is a risk acceptance that names the spec and the reason, or a proposed test. Risk acceptances show on the review card, and you decide whether the branch merges.
Set it up headless with tiden setup --api-token "$TIDEN_API_TOKEN" --workspace-id <id> --yes. tiden gate check prints the verdict per component and exits non-zero when the gate is blocked, not verified or waiting on a risk acceptance, so tiden gate check && deploy stops a release. Pass --format json wherever a script reads the output. The CLI never updates itself when CI is set.
Tiden checks the change against the intent captured before the session: the task, its promises and the specs in scope. If that intent was wrong, Tiden confirms you built the wrong thing. Whether the ask was right stays a human judgement, and the review card is where you make it.
Run tiden setup, then start your next agent session from a contract.