# Cross-Harness Export: One Component Pack Across Claude Code, Codex, and Ollama Cloud

Published: 2026-10-18 | Domain: https://getagentfoundry.si | Related Surface: https://proxy.baltor.ai

## Outcome first: write the capability pack once, run it in any harness

Baltor is not a vendor-locked model host. Customers bring their own API keys, sign-ins, or local endpoints (`127.0.0.1:11434`, Ollama Cloud, Claude Code, or Codex CLI) and use `getagentfoundry.si` to give their chosen harness verified materials, tools, and checks. Because the 8-file pack format uses plain Markdown, JSON Schema, YAML, and standard Python 3.11+ verifiers, a single pack mounts into all three harnesses without rewriting a line of logic.

## Mounting an 8-file pack across three harnesses

| Target Harness | Discovery & Rule Entrypoint | Action / Tool Binding | Independent Verification Hook |
| :--- | :--- | :--- | :--- |
| **Claude Code (`claude`)** | `CLAUDE.md` imports `AGENTS.md` + `RULES.md` | `.claude/settings.json` + `TOOLS.json` | Post-step hook runs `python3 verifier.py` |
| **OpenAI Codex (`codex`)** | `AGENTS.md` + `baltor_contract_card.json` | Sandboxed tool runner reads `TOOLS.json` | `WORKFLOW.yaml` gate executes `verifier.py` |
| **Ollama Cloud / Local** | System prompt loads `< 2 KB` contract card | OpenAI-compatible `tools` array from `TOOLS.json` | Loop Engine runner enforces `verifier.py` |

```bash
# Verify an 8-file AgentFoundry pack locally against its conformance kit
curl -fsSL https://getagentfoundry.si/feed.json -o /tmp/agentfoundry-feed.json
python3 -m json.tool /tmp/agentfoundry-feed.json | head -n 20
```

Whether a developer runs `gemma4:31b` on Ollama Cloud for overnight batch solving or Claude Code interactively on a workstation, the contract card and `verifier.py` guarantee identical structural invariants.