# Pome | AI safety infrastructure for agentic workflows

Machine-readable summary of https://pome.sh (human landing). This file is shown at https://pome.sh/?mode=machine and mirrored at /llms.txt.

---

## What Pome is

Pome is an AI safety company. We build the **self-healing staging layer for agentic workflows**.

As agents move from answering questions to taking actions (filing tickets, pushing code, charging cards, sending emails) failures become **wrong actions in production**, not bad answers. Pome sits between your agents and real services: tool use is exercised against **stateful digital twins** before anything reaches production.

**Tagline:** Self-healing staging layer for agentic workflows.

**One-liner:** Test your agents against digital twins of the APIs they call. Catch broken tool calls and hallucinated responses before users do.

**Primary CTA:** Get started → https://calendar.notion.so/meet/aofu/wj1bs4pbd (Request Closed Beta)

---

## Supported twins (stateful clones of real services)

| Twin    | Status |
|---------|--------|
| GitHub  | Live   |
| Stripe  | Live   |
| Zendesk | Beta   |
| Slack   | Beta   |
| Linear  | Beta   |
| Custom  | Talk to us → https://calendar.notion.so/meet/aofu/wj1bs4pbd |

Each twin models state transitions, side effects, and conditional behavior like the real API — not just HTTP record/replay.

---

## How it works — "See how your agents work before they hit production."

Run stateful simulations against multiple digital twins at every stage of development in an isolated sandbox. Test against edge cases that track real-time API changes and production failures to build confidence.

### 1. Simulation — Write test workflows
Describe the workflows your agent should complete and how it should complete them. Pome runs them against stateful digital twins.

### 2. Flight recorder — Watch the agentic "flight recorder"
Every tool call and state mutation is logged into a replayable audit trail. Rewind and debug multi-step failures that standard observability misses.

### 3. Guardrails — Disable destructive actions before production
Surface every destructive action from production traces. Toggle off unauthorized calls. Test scenarios inform future runs to prevent regressions (reversible: `GET listIssues`, `POST addComment` allowed; irreversible: `DELETE deleteRepo`, `POST pulls.merge` denied by default).

Flow: **AI agent → Pome → {GitHub, Stripe, Linear, Slack, Zendesk}**

---

## Scenarios — "Production-shaped runs you can replay."

Agents fail quietly: wrong tool, wrong assumption, wrong identity. Pome catches them against API twins before users do. Two example runs that didn't ship:

### GitHub Agent | Author Impostor Merge
A PR-review agent approves and merges based on a sloppy string match on the author handle: `ash_ketchum1` (look-alike, trailing digit) vs. the approved `ash_ketchum`. The merge tool fires. Production is one Slack post away from shipping an impostor's code.

**How Pome helps:** Run the agent against a digital twin of GitHub where the PR is authored by the look-alike `ash_ketchum1`. The criterion `state.pr.merged === false` fails the run in staging — Pome catches the bad `github.pulls.merge` before the impostor's code lands on main.

### Stripe Agent | Cross-system state drift (Refund Retry)
A customer-support agent picks up a Zendesk ticket requesting a $150 Stripe refund. The agent notices the charge already has an open Stripe chargeback for the same amount. But since a fraud check tool marked the customer as `low_risk`, it uses the tool incorrectly, and the agent approves a refund a second time.

**How Pome helps:** Run the agent against a digital twin of Stripe seeded with the open chargeback. The criterion `state.stripe.dispute.exists === false` fails the run in staging — Pome surfaces the broken `stripe.refunds.create` call on the trace before it ever becomes a $150 double-pay in production.

---

## Pricing — "Build reliably with Pome"

Stateful service twins and replayable audit logs, so every agent rollout ships with proof, not best effort. Four tiers. CTA for paid tiers points at the app sign-up unless noted.

| Tier   | Monthly | Annual               | Highlights |
|--------|---------|----------------------|------------|
| Free   | $0      | —                    | 3 concurrent isolated twins; 50 agent evals/mo; 10 audited runs/mo; 100 MCP/API calls/week; 3 months log retention |
| Hobby  | $19/mo  | $15/mo billed annually | 5 twins; 150 evals/mo; 20 audited runs/mo; 1 Million MCP/API calls/week; 6 months log retention |
| Pro    | $49/mo  | $39/mo billed annually | 10 twins; 300 evals/mo; 50 audited runs/mo; 2 Million MCP/API calls/week; 12 months log retention |
| Team   | Custom  | —                    | Everything in Pro; self-host; white-glove support; custom service clones; custom evaluation framework — contact via calendar |

**Platform comparison:**

- Concurrent isolated twins: 3 / 5 / 10 / Custom (Free → Team)
- Agent evals: 50 / 150 / 300 / Custom per month
- Audited runs: 10 / 20 / 50 / Custom per month
- MCP/API calls: 100/week (Free), 1 Million/week (Hobby), 2 Million/week (Pro), Custom (Team)
- Log retention: 3 / 6 / 12 months / Custom

**Team plan add-ons:** self-host option, white-glove support, custom service clones, custom evaluation framework.

**Annual billing:** save 20% on Hobby and Pro.

**Contact:** Need SSO, self-host, or a custom contract? https://calendar.notion.so/meet/gagandevagiri/zs9h2lxi

---

## Who Pome is for

Teams shipping AI agents that call real APIs (GitHub, Stripe, Zendesk, Slack, Linear, etc.) and need a **staging-grade** layer before production.

---

## URLs

- **Marketing:** https://pome.sh
- **Machine-readable view:** https://pome.sh/?mode=machine
- **This file (static):** https://pome.sh/llms.txt
- **Docs:** https://docs.pome.sh
- **Blog:** https://pome.sh/blog
- **Pricing:** https://pome.sh/#pricing
- **Book a demo:** https://calendar.notion.so/meet/aofu/wj1bs4pbd

---

## Company

- **Legal name:** Pome AI Ltd. (formerly Vakoi AI)
- **Founded:** 2024
- **Category:** AI safety, developer infrastructure, agentic AI
- **Stage:** Early access / waitlist
- **Social:** https://x.com/trypome | https://www.linkedin.com/company/pome-sh/ | https://github.com/pome-sh

---

## Key search terms

staging layer ai · digital twins for ai agents · agentic workflows · AI safety infrastructure · self-healing staging · agent evaluation · LLM tool-call testing · pre-production agent testing · agent observability · audit logs · MCP testing · flight recorder for agents · destructive action guardrails