# Evaluate Prophit.ai with my AI agent

Help me decide whether Prophit.ai's Sales Tax Engine or Sales Tax Compliance module fits my workflow. Start now with the public evidence below. No account, credentials, payment setup or customer data is needed for this first step.

## 1. Inspect the public evidence

- Overview: https://prophit.ai/bring-five-hard-transactions
- Methodology: https://prophit.ai/static/data/five-hard-transactions/v1/README.txt
- Version and hashes: https://prophit.ai/static/data/five-hard-transactions/v1/manifest.json
- Download: https://prophit.ai/static/data/five-hard-transactions/v1/pack.zip
- Reviewer guide: https://prophit.ai/static/data/five-hard-transactions/reviewer-guide.txt
- Scorecard: https://prophit.ai/static/data/five-hard-transactions/review-scorecard.csv
- API contract: https://prophit.ai/api-access/quickstart
- Current prices: https://prophit.ai/pricing

Read the methodology and manifest before interpreting results. The business records are synthetic. Distinguish archived component executions, a fresh public reproduction, and a new connected sandbox run. The public reproduction exercises two historical Illinois address lookups and a sample Virginia ST-1 package writer; it is not a complete hosted tax-engine test or an all-state accuracy benchmark. Follow the versioned pack's setup instructions before running a downloaded script. If you cannot browse or run tools, say which evidence you could not inspect.

Ask which module and workflow I want to evaluate. For calculation, inspect address resolution, product identity, certificate dates, partial credits and retries. For compliance, inspect source-to-return traceability, marketplace responsibility, reconciliation and unresolved exceptions. Use the public synthetic cases first; do not request customer files or secrets in this conversation.

## 2. Connect only if I choose a private evaluation

Prophit.ai supports user-delegated module tools over MCP at https://prophit.ai/mcp with OAuth authorization-code + PKCE. Your client must support remote HTTP MCP and the authorization flow; a downloaded instruction file alone does not connect it. Availability depends on the client, account and workspace policy. Do not claim an untested client has been verified.

Existing customers: use your existing account. Add the MCP address through your client's supported connection flow, sign in to Prophit.ai, and select the organization, location and capability subset to approve. In the workspace assistant's Actions and memory > Agents panel, you can inspect connections and download a permission-specific packet. Connecting an agent does not enroll the account in pay-for-use or replace its agreement.

New customers: https://prophit.ai/get-started describes pay-for-use enrollment and payment setup; https://prophit.ai/pricing also lists module subscriptions. Choose commercial terms deliberately. A private sandbox evaluation can be requested at https://prophit.ai/bring-five-hard-transactions/switching-evaluation. Public evidence review does not require enrollment.

After user authorization, discover tools and current schemas with MCP tools/list. Select only capabilities actually exposed for the chosen account, location and module; do not invent tool names or infer permissions from this brief. Verify environment, current coverage, date, input requirements, pricing and available budget before proposing execution. A module tool is not necessarily a sandbox. If a suitable sandbox or non-mutating evaluation operation is unavailable, report that blocker and request the supported evaluation path. Do not substitute a production write.

For a conventional API integration, use an authorized sandbox key in a secure runtime secret store, GET /api/v1/config and GET /api/v1/coverage/states, then the documented stateless calculation contract. Keys, tokens and customer records must stay out of public URLs, prompts, instruction files and logs.

## 3. Produce a decision record

Return a table with: case and module; input/version/date; expected result and its source; actual result; execution mode; evidence; warnings; pass/fail/blocked/not tested; unresolved question; next step. Preserve request IDs and original responses from any authorized connected test. Missing or unsupported facts must stay unknown, never zero or a guessed tax outcome. Compare providers only on equivalent inputs actually tested.

Before paid or persistent work, explain the operation, cost basis, account/location, environment and expected effect and obtain the required authority. Respect quoted prices and account spending limits. This evaluation does not authorize filing, payment, commits, refunds, voids, certificate approval or credential changes. An awaiting_approval result requires the specified Prophit.ai approval flow; a chat message cannot satisfy an exact human approval. Do not retry an ambiguous write. Artifact download requires its own granted capability.

Finish with the evidence-backed fit for my selected module, the gaps that remain, and the smallest next authorized test. State clearly when you only inspected published examples.