How it works

Discover. Decide. Prove. Turned on without touching the agent.

The engine behind Track, end to end — how it builds a sourced count of your agent estate, how it decides each action inline, how every decision becomes a checkable receipt, and how it's wired in with a one-line config change, not an SDK.

The core engine

Discover. Decide. Prove.

Find what's running, decide each action, prove what happened. Unlike posture- and identity-first tools, Track is built around the last two.

Every path — allowed or denied — ends the same way: a signed receipt.

Discover

Anchor
Systems you already trust — IdP, cloud, CI, repos.
↓
Reconcile
Against what's actually reaching a decision point.
↓
Show the gap
Governed, ungoverned, and the source of each.
↓
Keep watching
Re-check as new agents appear.
No install. Paths hashed, secret values never read.

Decide

Resolve
Who's asking, under what authority, on what.
↓
Apply policy
Your rules, plus the session's context.
↓
Decide
Allow, deny, restrict, or escalate to a person.
↓
Bind it
A one-time, payload-bound token — not advice.
Runs in your environment, inline on the action path.

Prove

Bind
Request, decision, execution — one hashed artifact.
↓
Sign
Every mediated action — allows too, not just blocks.
↓
Verify offline
Signature, chain, timestamp — re-derived with Track off.
↓
Keep it
Yours, even if you stop being a customer.
Tamper-evident, portable. Your keys throughout.
Getting it running

Turn it on without touching the agent.

No SDK. No code change. Just point the agent's MCP config at Track.

Same agent, same server — only the path between them changes.

Works with what you already run.
Claude Code, Copilot, Cursor, Codex, LangChain, CrewAI — any MCP client.
Your cloud. Your keys. Your rules.
Self-hosted, single-tenant. Nothing leaves your environment — ever.
Complements — doesn't replace.
Sits beside your IAM, SIEM, DLP. Governs the one thing they don't: the action.
Observe first, enforce when ready.
Read-only to start. Enforcement, one policy at a time.

Inline on the action path. Latency measured against your traffic, not a headline number.

See it decide on your own actions.

Pick one agent with real write access. We'll map it and show you the receipts.

Design a pilot