OpenAI terminal

Use an OpenAI-compatible model to inspect and reason about mail.

Both implementations are standalone agent processes. They use AgentMailer for identity, mailbox data, threads, drafts, and delivery; the example process owns the model call, execution state, and any external tool action.

They send untrusted content through an OpenAI-compatible model and keep external effects behind an explicit opt-in. The example is read-only.

Architecture boundary

AgentMailer is the communication substrate, not the workflow runtime. In a production deployment, persist event deduplication, checkpoints, business approvals, and tool results in your own application. Use AgentMailer message and thread IDs to correlate that state, and stable idempotency keys for email writes.

Environment

1
export AGENTMAILER_API_KEY="..."
2
export AGENTMAILER_INBOX_ID="inbox_..."
3
export LLM_API_KEY="..."
4
# Defaults to Fireworks; override for another OpenAI-compatible provider.
5
export LLM_BASE_URL="https://api.fireworks.ai/inference/v1"
6
export LLM_MODEL="accounts/fireworks/models/llama-v3p3-70b-instruct"

TypeScript

1
pnpm --filter @agentmailer/example-openai-terminal start

Python

1
cd examples/agentmailer-openai-terminal/python
2
python -m venv .venv && source .venv/bin/activate
3
pip install -r requirements.txt
4
python agent.py

Use a dedicated test inbox. Treat model output as a proposal, record idempotency keys in your application before side effects, and require human review for consequential actions.