Use judgment where it matters. Use tools where they suffice. Delegate when it helps. Verify what comes back.
Hephaestus is an instructional plugin for efficient AI execution. It teaches an agent how to work, rather than introducing another orchestration platform.
Designed for ChatGPT web Chat and Work, with capability-based behavior in other compatible runtimes. Installation and runtime support depend on the account and workspace; a valid package is not proof of successful cloud installation.
After installation, select Hephaestus or ask:
Use Hephaestus to complete this task efficiently. Delegate independent work to available subagents when worthwhile, and verify the result.
The core method needs no terminal, API key, MCP server, Fleet connection, or local repository. The actual task may still need authorized tools or source access.
- Retrieve the relevant slice of context instead of repeatedly loading whole histories.
- Use deterministic operations before model inference when appropriate.
- Give workers bounded assignments and request compact results with evidence.
- Select an economical model only when the actual interface and policy permit it.
- Verify acceptance criteria, recover from failed attempts, and preserve useful state.
Without worker tools, the agent applies the method directly. It never invents a subagent or claims an unconfirmed model choice. Smaller parent context is not automatically lower total cost.
This repository includes a one-plugin marketplace at .agents/plugins/marketplace.json.
For an eligible workspace administrator, the import source is https://github.com/eidos-agi/hephaestus, Path is empty, and the default branch is main. Importing a repository is a separate step from committing its files. See cloud installation for exact settings and account-dependent alternatives.
No mcp.json, .mcp.json, app dependency, hook, or runtime installation script is declared. This deliberately avoids the documented MCP-related Desktop only restriction for imported plugins. Existing authorized tools remain optional execution routes, not plugin dependencies.
The live MCP at https://hephaestus.eidosagi.com/mcp serves current, versioned guidance through three read-only tools. Connect it separately in ChatGPT using OAuth; a personal API token is entered once on Hephaestus’s connection page. The instruction package remains usable without that connection.
See live guidance and administration and deployment verification. The Worker and its development dependencies are excluded from the installable instruction archives.
plugin.json Portable Agent Plugins entry point
.codex-plugin/plugin.json Legacy compatibility metadata
.agents/plugins/marketplace.json GitHub workspace-import catalog
skills/hephaestus/SKILL.md Small, cloud-readable entry skill
skills/hephaestus/references/ Selectively loaded recipes and evidence rules
The root manifest is canonical. OpenAI presentation metadata lives in extensions.com.openai; the legacy manifest mirrors it for older consumers. Modern clients do not merge the two overlays.
These are maintainer commands, not runtime prerequisites. They use only Python's standard library:
python3 -m unittest discover -s tests -v
python3 scripts/package.pyThe packager produces dist/hephaestus-plugin-0.1.0.zip and dist/hephaestus-skill-0.1.0.zip with fixed timestamps, an explicit file allowlist, and SHA-256 receipts. The skill ZIP has a top-level hephaestus/ directory containing SKILL.md and references. Import acceptance must still be tested in the target surface.
See evaluation for fresh-chat scenarios, cost accounting, and unverified release gates. Static tests do not certify model behavior, token savings, or ChatGPT installation.
Current documentation reviewed September 12, 2026:
- OpenAI plugin packaging
- OpenAI skills authoring
- OpenAI subagent behavior
- ChatGPT plugins, permissions, and surface restrictions
- GitHub marketplace import
Hephaestus v0.1.0 is an initial implementation, not a public-directory listing or a measured savings claim.