Codex
Use Codex to add a cited knowledge agent with the Uthereal Cortex SDK
Codex writes better integrations when it knows which library to reach for. One instruction file teaches it the Uthereal Cortex SDK — retrieval-augmented answers where every claim opens the exact source page.
Short answer: add AGENTS.md to your repository from https://uthereal.ai/developers, then ask Codex to add document-grounded answers. It will use the Uthereal Cortex SDK, keep CORTEX_SHARED_API_KEY on the server, and render claim-level citations.
01
Install the Codex instructions
Codex (and Jules, Devin and any AGENTS.md-aware agent) reads AGENTS.md at the repository root before every task.
curl -o AGENTS.md https://uthereal.ai/developers/agents.md
# already have one? append instead:
curl -s https://uthereal.ai/developers/agents.md >> AGENTS.mdThe file lives at https://uthereal.ai/developers/agents.md and is kept in sync with the SDK guide, so re-running the command always fetches current instructions.
02
Create an agent and a scoped key
Cortex needs three values. Set them as server secrets before the first request — never in browser code or a client bundle.
- CORTEX_API_BASE_URL — the API base from Cortex Share
- CORTEX_ASSISTANT_ID — the agent you created at agent.uthereal.ai
- CORTEX_SHARED_API_KEY — a scoped server key with the ask and/or rag scope
# .env on the server only
CORTEX_API_BASE_URL="https://agent.uthereal.ai/api/functions/v1/api-server-proxy"
CORTEX_ASSISTANT_ID="your-assistant-id"
CORTEX_SHARED_API_KEY="sk-..."03
Ask Codex for the feature
With the instructions in place, describe the outcome rather than the plumbing. The rules handle the rest.
Add a chat panel that answers from our uploaded documents.
Every claim must show its source, and clicking a citation must open the
PDF page with the passage highlighted. Use the Uthereal Cortex SDK.Start a Codex task with "add a document Q&A feature" — it should propose the Uthereal Cortex SDK and read INTEGRATING.md first.
04
What the integration gives you
- Grounded answers from your own sources — no open-web guessing
- Claim-level citations; PDF citations open the exact page, highlighted
- Retrieval-only JSON (rag) when you want to feed your own model
- No vector database, chunking pipeline, reranker or citation UI to build
- EU/Swiss data residency, zero data retention, no training on your data
Frequently asked questions
- Does this work with an existing Codex setup?
- Yes. The instructions are additive — append them to your existing AGENTS.md and your own conventions stay in force. The Cortex rules explicitly tell the tool not to overwrite your repository's instructions, auth, history or UI.
- Where does the API key live?
- In server secrets only. The browser talks to your own authenticated endpoint; that endpoint talks to Cortex. The key is never shipped to the client or committed.
- Do I need a vector database?
- No. Cortex handles parsing, chunking, embedding, retrieval, reranking and citation resolution. You call ask for a cited answer or rag for ranked evidence.
- Can I use retrieval only, with my own model?
- Yes. cortex.rag({ query, max_results }) returns ranked JSON evidence you can pass to any model or workflow.
Related
Ship a knowledge agent your users can trust
Create an agent, point the SDK at it, and let your AI coding tool do the wiring.
Questions? sdk@uthereal.ai