Skip to content

Enterprise deploymentFully managed today; customer-hosted available as a pilot.Contact salesView pricing

+Coding agents

Coding agents, governed without the friction.

One command connects Claude Code, Cursor, Codex, Copilot CLI and OpenCode to Intertrace. Every tool call is decided before it runs: routine work flows, risky actions ask, and the dangerous ones never run.
Install and connect
curl -fsSL https://platform.intertrace.ai/install.sh | sh
itr connect     # every coding agent on this machine
itr doctor      # runs each installed hook the way its editor does

+Also

npm install -g intertrace on any OS with Node 18+, or irm https://platform.intertrace.ai/install.ps1 | iex on Windows. For a hook developers can't edit, itr connect claude-code --managed writes it to Claude Code's managed settings; for the other editors, deploy it with your MDM or the editor's own policy files.

Setup guide →

+Try it

What would Intertrace do?

Type any command a coding agent might run. This is the same reading of the command that the enforcer makes, applied with the default Coding role.

What Intertrace reads

It force-pushes and changes a protected branch (main).

Consequence: Changes that can't be undoneproduction

What happens

Asks a person first

The call waits while a person decides in Intertrace, notified in Slack, Microsoft Teams or email. Approved, it runs with exactly these arguments; denied, it never does.

A preview with the default Coding role in development, using the same reading the enforcer uses. Your organization's rules, guardrails and approvals can add to it, and a new agent asks more often during its learning period. The live decision is always the enforcer's.

+How it works

The same decision as every other agent.

01

Connect in one command

itr finds the coding agents on the machine and installs a pre-tool hook in each, touching only its own entry. Sign-in mints a device key bound to your organization's agent for that tool.

02

Every call is decided by the enforcer

Each tool call goes to the same enforcer as your production agents: roles, consequence tiers, guardrails, approvals and floors. It reads the arguments, so a force push to main isn't treated like a push to a branch.

03

Ask only when it matters

Routine development work runs. Things that only touch the developer's own machine ask the developer in the editor. Anything that leaves the company, moves money or reaches production waits for a person you choose.

04

Proof of what ran

Post-tool reports join each decision to what actually executed. Every call lands in the activity ledger and the conversation's trace, next to every other agent you run.

+Every editor, honestly

What each coding agent lets us guarantee.

Editors don't all expose the same hooks. This is what each one supports today. Where an editor can't show the developer a question, the ask waits for a person instead of running.

Coding agentDecided before it runsDeveloper answers in the editorRecords what actually ranBlocks if the hook itself fails
Claude Code
✓Every tool, before it runs
✓Interactive sessions
✓Succeeded and failed calls
✓A broken hook blocks
Cursor
✓Shell, MCP and file reads
PartlyShell and MCP
✓Shell, MCP and edits
✓A broken hook blocks
Codex CLI
✓Once you trust the hook in /hooks. itr can't confirm that trust, so Codex is never shown as Protected
—Asks wait for a person
✓Every executed call
—Codex runs the call if its hook can't start
GitHub Copilot CLI
✓Every tool, before it runs
—Asks wait for a person
—Not wired yet
PartlyCrashes block; a host timeout doesn't
OpenCode
✓Every tool, before it runs
—Asks wait for a person
✓Every executed call
✓The plugin blocks unless allowed

On every editor, anything itr can see going wrong (no key, Intertrace unreachable, an undecided or expired approval) is a deny. Only workspace reads may continue offline, and that allowance is explicit and logged. Windsurf and Gemini CLI have no pre-tool hook: itr routes their remote MCP servers through Intertrace (local command servers aren't routed). For Devin, itr prints the MCP settings to add by hand.

+Connected is not protected

We only say protected when we've proved it.

Verified, not assumed

itr doctor runs the installed hook the way each editor runs it and sends a real probe through Intertrace. “Protected” appears only after that probe comes back. Codex never shows Protected, because itr can't confirm that you trusted its hook.

The agent can't switch it off

Edits to its own hook settings, turning hooks off, or starting another agent with its safeguards disabled are blocked before they run.

If the hook is removed

Hooks live in the developer's own settings, so a person can remove them. While the post-tool report still runs, a tool that ran with no decision opens a high-severity finding. To keep the hook in place, deploy it through managed settings.

+Built for developers

Low friction is a feature, not a setting.

+Runs without asking

  • Reading, searching and editing code
  • Tests, builds and installs in development
  • Commits and pushes to the agent's own branches
  • Deletes in a development workspace

+Asks first

  • Pushing or merging to main and protected branches
  • Anything that reaches production
  • Sending data or messages outside the company
  • Running downloaded installers or reading a project's .env (the developer can answer)

An approval covers exactly the call that asked for it: the same arguments, the same agent, the same session. Changing the command opens a new one. Your organization's rules, roles and guardrails decide what “asks” means for each team.

+ Coding agents

Govern every coding agent in one command.

Book a walkthrough, or connect your own editor and watch the first decision land.