agent.run()Marketplace

List your template

The marketplace is one file, .agentrun/marketplace.json in Parcha-ai/agentrun-templates. Your template stays in your repository. Listing it is a pull request that adds one line.

Paste this into Claude Code, Codex or another coding agent, in the repository that holds your template. It checks the template, fixes the gaps and opens the pull request.

Help me list my AgentRun template on the AgentRun template marketplace (https://marketplace.agentrun.ai/).

Work in the repository I have open. Ask me before you push, tag or open a pull request. Never invent an eval result or a sample output: run the template to get them.

Reference: https://marketplace.agentrun.ai/list-your-template.md

1. Find the template folder: the one that holds agent.md. If there are several, ask me which one.
2. Check it against the bar below. Tell me each gap in one line, then fix the ones I approve.
3. Run `npx @parcha/agentrun template validate <folder>` and fix what it reports until it passes.
4. Check the GitHub repository is public. If it is private, stop and tell me: private repositories can't be installed.
5. Pin a version: propose a tag (for example v1.0.0) on the commit that passed, and create and push it once I agree.
6. Write the marketplace entry and show it to me. Then fork Parcha-ai/agentrun-templates, add the entry to .agentrun/marketplace.json, add the sample run under samples/<template>/<name>/, and open a pull request.
7. Give me the pull request link and say what the reviewers will check.

The bar:
- Typed inputs for what each run is about, and parameters only for setup (a schedule, a channel).
- One or two skills that hold the method, each SKILL.md under 200 lines.
- An output schema (JSON Schema 2020-12) with a report_markdown string marked contentMediaType text/markdown, beside the typed fields.
- A per-run budget (budget.per_run_usd), and learning: true when it runs on a schedule.
- At least four eval cases: one prompt injection inside a fetched page, one where there is nothing to find and it must not invent anything, and one per thing the description promises. A scheduled template adds a no-repeats case.
- A sample run in samples/<template>/<name>/ of this repository: input.json and an output.json that matches the schema.
- A summary of 20 to 140 characters in the index entry that says what it does in plain words, and a description in agent.md with the detail.

The entry:
{ "name": "your-template", "source": "github:you/your-repo/path/to/template@v1.0.0", "category": "research", "tags": ["your-tag"],
  "summary": "One line, what it does, shown on its card and under its title.",
  "setup": "Optional. What to do before installing, shown once in Install.",
  "guide": "Optional. An https link to a step-by-step setup guide, linked under setup." }

Categories: engineering, security, sales, marketing, research, finance, operations, compliance, personal. Up to 8 tags, lowercase letters, digits and dashes. Pin the source with @ref (a tag or a commit).
  1. Put the template in a public GitHub repository. One folder: agent.md with its front matter, skills/, evals/ and the output schema. Private repositories can't be installed.
  2. Check it. Run agentrun template validate <folder>. It runs the same checks AgentRun runs on install.
  3. Open a pull request that adds one entry. On Parcha-ai/agentrun-templates, add your template to .agentrun/marketplace.json. Pin the source to a tag or a commit with @ref: the page and the install link use exactly that version.
  4. We review it. We read the instructions and skills, run the evals, and check it meets the bar below. When it merges, the site rebuilds and your page goes up.
  5. Ship a new version. Open another pull request that moves the @ref. Installs already made keep the commit they were installed from.

The entry:

{ "name": "your-template", "source": "github:you/your-repo/path/to/template@v1.0.0", "category": "research", "tags": ["your-tag"],
  "summary": "One line, what it does, shown on its card and under its title.",
  "setup": "Optional. What to do before installing, shown once in Install.",
  "guide": "Optional. An https link to a step-by-step setup guide, linked under setup." }

Categories: engineering, security, sales, marketing, research, finance, operations, compliance, personal. Up to 8 tags, lowercase letters, digits and dashes.

The same one the bundled templates meet.

Add a measured.json beside agent.md once you have run your evals, and the page shows how many cases pass on each harness:

{ "runs": 12, "model": "openrouter/particle-ai/particle-1", "harness": "pi", "measured_on": "2026-11-14", "passed": { "pi": "5/5", "claude-agent": "4/5", "codex": "5/5" } }

What a template requires (its integrations, web search, a schedule) comes from its agent.md: its tools, MCP toolkits and triggers. There is nothing to write for it.

A template that searches or reads the web adds "live": { "cases": 1, "passed": { "pi": "2/2" } }: its cases named live-*, which give it no pages and run on the real web. passed counts the other cases. Without live, the page says it has not been tried on the live web.