frumu-ai/tandem — reverse-engineered prompt

Build me a local first app called Tandem that acts like a control layer between AI agents and real company tools. I want agents to be able to reason and draft work, but Tandem should decide what they’re allowed to see, which tools they can use at each step, what company memory they can access, and which actions need a human approval before anything actually happens.

Please make it support durable workflow runs, not just chat transcripts, with a clear audit trail that records tool activity, approvals, artifacts, and decisions outside the model context. It should work for common cases like letting an agent draft an email or prepare a repo change, but not send or merge anything without approval. I also want tenant and workspace boundaries so access stays scoped correctly.

It should be usable locally, headless, and through a simple control interface, and it should be able to work with different model providers like OpenAI, Anthropic, OpenRouter, or local Ollama. Look up current docs if you need to.