Using OpenCode or Pi?

From coding to
checked delivery.

Put the requirements, implementation, and checks in one workflow. Judge quality by what works—and speed by time to an accepted result.

A harness that strategizes and validates before it acts.

Try it on one issue See the evidence

Use your own provider. No Cloud subscription required.

What Carthagent adds to the workflow

  1. Define done.

    Declare acceptance criteria and map them to executable checks before implementation.

  2. Keep proof current.

    Run real checks. Relevant source changes invalidate earlier evidence; passing an old snapshot does not clear the new one.

  3. Check the handoff.

    Bounded repair addresses failed checks. Finish as verified only with current required evidence; otherwise, blocked.

“Verified” means the declared required checks passed with current evidence. Incomplete checks can still miss bugs or requirements. Human judgment remains necessary.

Inspect the workflow and its limits

A real Pi comparison, including the failures.

Same voxel-world task. Same model: DeepSeek v4.1 Flash.

Carthagent coding run
8m 58s
Pi coding run
12m 00s
Playable acceptance
Neither passed

One attempt per tool, with different workflow conditions. These are coding-run times, not time to an accepted result. This experiment establishes neither a general speed advantage nor higher quality; there is no OpenCode result here.

Read the results, captures, and independent checks

Give it your next issue. Keep your provider.

Choose a bounded change in an existing project. Define acceptance before the run, then inspect the diff and exercise the result yourself. Compare rework and elapsed time with your usual workflow.

Install Carthagent Ship: hosted models

Cloud is optional: no API-key setup. Builder is $1 for the first month for new customers, then $10/mo. Pro starts at $25/mo; Max at $50/mo. Usage is metered.

Official site: carthagent.xyz. CLI: ctg. Source: B-icy/Carthagent on GitHub. Not affiliated with carthagent.com.