Comparison · Coding agents
Curiosity and OpenAI Codex
Different jobs, used together
Codex writes the code, locally or as parallel cloud tasks; Curiosity is the platform the tool runs on.
Your data · your environment
Who each tool is for
Different problems, honestly stated
Curiosity
For industrial and regulated teams whose knowledge sits in engineering systems and structured records, where the answer depends on how those records connect, and where the deployment has to run on their own infrastructure.
OpenAI Codex
OpenAI’s coding agent: an open source CLI, an IDE extension, a coding view in the ChatGPT apps, and Codex Cloud, which runs tasks in environments OpenAI manages and opens pull requests. Locally it works inside an OS sandbox with network access off by default. It suits teams on ChatGPT plans or the OpenAI API who want parallel cloud tasks and CI automation.
How Curiosity compares
Where the boundary falls
Codex is not an alternative to Curiosity, and teams use both. It runs many tasks at once and works safely in a sandbox. What it does not bring is the system an enterprise tool runs on: connected records, permissions per record, sign in, an audit trail and hosting. Curiosity is that system, and through curio Codex builds on it.
The platform the code runs on
A coding agent writes code. An enterprise app also needs its data connected and kept current, a data model, search, sign in and hosting. Curiosity ships those, so the agent writes the part that is yours.
Permissions the code cannot forget
Code an agent writes checks access only where it remembers to. In Curiosity, access is synced from the source systems and applied per record, so a tool built on it inherits what each user may see.
A change a person approves
Through curio the agent works in Sudo’s sandbox: it edits, builds against the workspace and stages a commit. Nothing runs until a person approves the diff.
Keep the agent you have
curio is a command line, so any coding agent that can run a shell command can work on your workspace. Your developers keep their tools; the platform keeps its rules.
Out of the box
What an enterprise app needs before its first screen
OpenAI Codex can write each of these. Then your team owns it: keeps the connectors current as the source systems change, and shows an auditor that every screen checks access. In Curiosity they come with the platform.
Data from your systems
70+ systems connected, kept in sync
Connectors the agent writes and your team runs
A data model
A typed graph of records and their links
A schema the agent designs for each app
Search
Keyword, semantic and hybrid search in one query
A search stack to choose, host and tune
Permissions
Synced from each source, applied per record
Access checks the code has to remember
Sign in
SSO through Google, Microsoft, Okta, Auth0 or SAML2
An identity integration for each app
Audit trail
User actions, downloads and AI interactions logged
Logging the agent adds when someone asks
Change control
Every change a diff a person approves
Your repository’s pull request review
Hosting
Your servers or private cloud, with replicas and backups
Infrastructure your team builds and runs
Use them together · curio
OpenAI Codex writes it. Sudo checks it. You approve it.
curio is Curiosity’s command line. It gives OpenAI Codex a session on your workspace: Sudo’s skills to read, a sandbox to edit in, and a build that checks the change against the workspace. The agent never gets more than that session.
- 01
Install curio
One command on the machine where Codex runs, then sign in to your workspace in the browser.
- 02
Give OpenAI Codex the skill
curio skills install ./.agents/skillsputs the skill where Codex looks for skills in a repository. - 03
Approve what it builds
The agent edits and builds in a session, and stages a commit. A person approves it on Sudo’s review screen, and only then does the workspace run it.
curl -fsSL https://curiosity.sh/install.sh | bash
curio login --server https://studio.example.com
curio skills install ./.agents/skillsIn Codex Cloud, allow your workspace’s domain in the environment’s internet access, and set CURIOSITY_SERVER and CURIOSITY_TOKEN for curio. Without the approve right, the default, approval stays with a person.
Side by side
Feature comparison
The same five questions, answered for both.
Deployment
Cloud or self-hosted, ready to use
Your machine or OpenAI cloud environments
Data model
Managed context layer and knowledge graph
Files in your repository
Industrial fit
Built for industrial and enterprise teams
Developers and software teams
Data reach
70+ systems connected, and yours through a connector you write
Local files, shell, MCP servers, web search
Residency
EU hosted, or entirely your own infrastructure
OpenAI by default, or a local model
Customization
What your team can change
Your own code, interface, data model and configuration. OpenAI Codex as its own documentation described it in October 2026. Where it says Not available, the vendor documents no such capability.
Your own code
C# endpoints, AI tools and tasks, compiled in Studio
Writes and runs any code your shell runs
Your own interface
Your own front end in C#, deployed from git
Writes interface code, does not host it
Data model
Typed node and edge schemas you define
Not available, works on your files
Configuration as code
Every change committed to git, pushed to your remote
AGENTS.md, config.toml, skills, hooks, rules in git
APIs and SDKs
REST API, C# and Python library, connector SDK
Codex SDK, app server, codex exec, GitHub Action
AI integration
What AI can build and reach
Whether AI can build on it, who reviews the result, and which models it runs. OpenAI Codex as its own documentation described it in October 2026.
AI that builds
Sudo drafts endpoints, tools, schemas and screens
Plans, edits files, runs commands, opens pull requests
Review before it applies
Every change a diff a person approves, in a sandbox
Sandbox by default, approval prompts, automatic review option
Model choice
Any major provider, or an OpenAI compatible endpoint
OpenAI models, or a provider with a Responses API
Models on your servers
Open weights on your servers, or air gapped
Local models through Ollama or LM Studio
Assistant permissions
Built in tools search as the user, per record
Your credentials, inside an OS sandbox by default
Agent protocols
MCP server and an OpenAI compatible API
MCP client, app server over JSON-RPC
Which to use
When each is the right choice
Building on Curiosity does not mean giving up OpenAI Codex. The question is what it builds on.
OpenAI Codex alone
- You want tasks run in parallel against your repository.
- The application already has its data, sign in and hosting.
- The tool reads one system, or none.
- It is a prototype, not a tool people run on production records.
OpenAI Codex on Curiosity
- The tool needs records from several systems, kept current.
- Each user may see only what the source systems show them, per record.
- It runs on your infrastructure, with SSO and an audit trail.
- You want every change built, reviewed and approved before it runs.
At Airbus, we developed the next generation of customer support AI assistance tools with Curiosity, enhancing our technical support’s ability to access and navigate vast data from multiple sources, including millions of documents.
Services Innovation Team
Airline Services at Airbus
Coding agents
Other comparisons in this group
Claude Code
Claude Code writes the code from your terminal; Curiosity is the platform the tool runs on.
CompareGitHub Copilot
Copilot writes the code and opens the pull request; Curiosity is the platform the tool runs on.
CompareGoogle Antigravity
Antigravity runs the agents that write the code; Curiosity is the platform the tool runs on.
CompareYour data. Your infrastructure.