What Omnodex captures, and where it stops
Omnodex records agent activity through each platform's own hooks, or through an MCP proxy where hooks are not available. What we see on a platform is what that platform reports. This page lists both, as of the current release.
Hooks
The agent calls Omnodex at defined points, such as before and after a tool runs. Hooks run outside the model context, so they add no tokens, and they see built-in tools such as the shell and file edits.
MCP proxy
The Omnodex plugins add a local MCP proxy that sits between the agent and its MCP servers and records each call routed through it. Use it alongside hooks or instead of them. Where a call is seen by both, the dashboard pairs the two records rather than counting the call twice.
| Platform | How we capture | What we see | Limits |
|---|---|---|---|
| Claude Code CLI and IDE extensions | Hooks | Sessions, tool calls before and after they run, failed calls, file reads and writes, prompts, subagents, permission requests and denials | Only what Claude Code reports through hooks. The model's own reasoning and replies are not captured. |
| OpenAI Codex Codex CLI and the Codex mode in ChatGPT Desktop | Hooks, plus the MCP proxy through the Codex plugin | Sessions, shell commands, file edits, MCP calls and local function calls | Hosted tools do not pass through local hooks. Codex does not reliably report failed calls, so failures can appear without an outcome. Prompts and subagents are not recorded yet. |
| Google Antigravity CLI, Desktop and IDE | Hooks, plus the MCP proxy through the Antigravity plugin | Sessions and tool calls before and after they run | Antigravity does not include tool responses in its hooks. |
| Cowork Desktop Claude desktop app | MCP proxy through the Cowork plugin | Calls to the MCP servers routed through the proxy | Built-in tools that do not go through an MCP server are not seen. |
Claude Code
CLI and IDE extensions
How we capture
Hooks
What we see
Sessions, tool calls before and after they run, failed calls, file reads and writes, prompts, subagents, permission requests and denials
Limits
Only what Claude Code reports through hooks. The model's own reasoning and replies are not captured.
OpenAI Codex
Codex CLI and the Codex mode in ChatGPT Desktop
How we capture
Hooks, plus the MCP proxy through the Codex plugin
What we see
Sessions, shell commands, file edits, MCP calls and local function calls
Limits
Hosted tools do not pass through local hooks. Codex does not reliably report failed calls, so failures can appear without an outcome. Prompts and subagents are not recorded yet.
Google Antigravity
CLI, Desktop and IDE
How we capture
Hooks, plus the MCP proxy through the Antigravity plugin
What we see
Sessions and tool calls before and after they run
Limits
Antigravity does not include tool responses in its hooks.
Cowork Desktop
Claude desktop app
How we capture
MCP proxy through the Cowork plugin
What we see
Calls to the MCP servers routed through the proxy
Limits
Built-in tools that do not go through an MCP server are not seen.
Capture depends on each platform's hook system, which can change between versions. Setup steps for every platform are in the documentation.
Not covered yet
- GitHub Copilot
- ChatGPT Chat and Work modes
- Cursor
- Browser chat on claude.ai or chatgpt.com
- Mobile apps
- Cowork tasks that run in the cloud (coming soon on Pro)
What capture does not mean
- Recording an action is not the same as stopping it. Omnodex observes and flags today; blocking is on the roadmap.
- The local event log is append-only, which keeps a complete history, but it is not tamper-proof against someone with access to your machine.
- A platform that is missing here, or a call that does not pass through a hook or the proxy, leaves no record.
Using a platform we do not cover? Tell us.