Spend caps, vendor allowlists,
and audit trails for your coding agents.
One config line. Sixty seconds. Your next Claude Code, Cursor, or Codex session is governed.
npx yelt initInstall path is the product.
- 0:00Land on yelt.ai/dev, run npx yelt init
- 0:15Yelt detects Claude Code, Cursor, Codex configs
- 0:30Magic-link auth in browser, no password
- 0:45Five default policies install · workspace ready
- 1:00Next agent action is governed and audited
Policy as instruction.
Policy as enforcement.
Every install drops a SKILL.md into ~/.claude/skills/yelt/ (and the equivalent for every other detected host) so the agent reads the rules as instructions before it acts. The gateway then enforces them at runtime — even when the model forgets.
- Plain-English summary
- Hot-reloads on policy change
- One file per agent host
- Removable with yelt uninstall
# Financial Operations Policy This skill applies whenever you are about to perform a financial action. ## Rules - Payments above $500 to a new vendor require human approval. - Refunds above $100 require approval. - Never send to OFAC-sanctioned addresses. - Customer PII must be redacted in audit. ## How to comply - Route all financial tool calls through the Yelt MCP gateway at mcp.yelt.ai. - If a call is held for approval, wait — do not retry with reduced parameters to slip under a limit. # version 47 · sarah@acme.com · 2026-04-22 14:00 UTC
Auto-detects every host
Claude Code at ~/.config/claude. Cursor at ~/.cursor. Codex CLI. Cline. Factory Droid. We sniff the configs so you don't pick.
Skill files install too
Yelt writes a SKILL.md into each detected host so the agent reads the rules as instructions. Belt and suspenders.
Promote to org without re-onboarding
Started solo? When your team adopts, click upgrade and your personal policies migrate to the org's defaults.
Hosted gateway by default
Dev tier routes through mcp.yelt.ai — no sidecar to run, no Docker daemon to keep alive. Self-hosted is enterprise.