Target platforms

Pick one or both. Claude emits .claude/agents, a skill and optional CLAUDE.md; Codex emits AGENTS.md and .codex config layers.

Lowercase slug: letters, digits and hyphens, 2–64 characters. It becomes the file name.

Least privilege by default. Full is capped: Claude never gets bypassPermissions and Codex stays inside workspace-write.

One sentence, up to 200 characters.

What the agent must achieve, up to 2,000 characters.

Optional. Used as the Claude description and the Codex role description; defaults to the role.

1/10

Concrete duties, up to 10 items of 300 characters each.

1/10

What the agent must not do, up to 10 items.

1/10

Three mandatory safety rules are always added.

1/10

Things the agent must refuse, up to 10 items.

Allowed tools

Leave all unchecked to get the read-only default (read, grep, glob) plus a warning.

Optional, up to 2,000 characters. Copied into the generated instructions.

Fill in the name, role and goal, then generate to see the files here.

What is the Agent Generator?

The Agent Generator turns a short description of an AI agent into the configuration files that Claude Code and OpenAI Codex actually read. Instead of hand-writing YAML frontmatter, you choose a platform, a name, a role and a goal, then add responsibilities, non-goals, guardrails, refusals, an allowlist of tools, a model tier and an optional MCP server. The generator renders a Claude subagent definition, a skill file, an AGENTS.md project document and Codex config layers from the same input.

Everything is deterministic and local to the request: there is no model call and no stored prompt, so the same input always produces byte-identical files. Generated text is escaped for YAML, TOML, JSON and Markdown, secrets are redacted, and warnings point out account-dependent model ids, placeholder tokens and oversized AGENTS.md files. Download the tree as a ZIP and commit it to your repository.

How to generate an agent

  1. Pick Claude Code, Codex or both.
  2. Enter a lowercase agent name, a one-sentence role and the goal.
  3. Add responsibilities, non-goals, guardrails and refusals, one item per row.
  4. Choose the tools, model tier and language, then generate.
  5. Review the warnings, copy individual files or download the whole tree as a ZIP.

Frequently asked questions

Does the generator call an LLM?

No. It is a pure template renderer: the generator makes no model call and no outbound request, and it stores nothing. The shared auth and rate-limit layer may use Supabase for sessions and quotas.

Which files are produced for each platform?

Claude Code gets .claude/agents/<name>.md, .claude/skills/<name>/SKILL.md, optional CLAUDE.md and optional .mcp.json. Codex gets AGENTS.md, .codex/config.toml and .codex/agents/<name>.config.toml. Large AGENTS.md files are split at line boundaries and reported in the warnings.

Can I use the generated configuration commercially?

Yes. The files are plain configuration you own; replace the Arttron MCP placeholder token with your own key before committing, and verify the account-dependent Codex model id.

Are secrets sent to the server?

Do not put real secrets in the form. Secret-like values are redacted before rendering, but the safest rule is to use placeholders and environment variables.

The generator runs on our server and makes no LLM call or outbound request; your input is not stored. The shared auth and rate-limit layer may use Supabase for sessions and quotas. Copy the files or download the ZIP.

Back to home