Key facts
| Endpoint | POST https://api.plugsky.com/v1/chat/completions with streaming and function calling |
| Compatibility | Same shape as OpenAI; change base_url and model name |
| Integration | Your service layer connects the API to the helpdesk; ticket state stays in the helpdesk |
| Models | 30+ models; route simple intents to cheap aliases and complex cases upward |
| Pricing model | Flat monthly self-serve plans with unlimited fair-use usage |
| Free tier | Free plan with 2 free AI models, no card required |
| Governance | Scoped keys per channel, audit logs and a defined PII policy |
| Roadmap | Moderation and audio endpoints are coming soon |
TL;DR
- Start with suggested replies for agents before auto-sending anything.
- Stream into the agent console; the helpdesk remains the system of record.
- Redact PII before prompts where policy requires it, and log every turn.
- Route by intent: cheap aliases for FAQs, stronger ones for complex tickets.
- Measure suggestion acceptance and handling time, not just deflection.
How it works, step by step
- Change the base URL to Plugsky and confirm an existing support script still streams correctly.
- Build a service layer between the helpdesk and the API that assembles ticket context per turn.
- Start with suggested replies: the model drafts, the agent edits and sends.
- Add tools for account lookups where the helpdesk already has permissioned data access.
- Log every prompt, response, edit and send decision for QA and evaluation.
- Graduate selected intents to auto-send only after accuracy is stable in review.
Original data
Try it yourself
Open the streaming API tester →
Architecture around your helpdesk
The API is a component, not the platform:
- Helpdesk: remains the system of record for tickets, customers and history.
- Service layer: assembles context, applies policy, calls the API and writes results back.
- Suggested reply service: streams a draft to the agent console with sources and confidence.
- Auto-send service: limited to approved intents with strict guards and logging.
- Analytics: tracks acceptance, edits, handling time and escalations per intent.
Implementation details
Support integration is mostly hygiene:
- Assemble context per turn: ticket subject, recent messages, customer plan and any retrieved policy text.
- Apply a PII policy before the model call — redact or tokenize identifiers the model does not need.
- Use streaming to reduce perceived latency in the agent console.
- Expose tools only where the helpdesk already enforces permissions, and never bypass them.
- Keep prompt versions pinned and stored with each response so quality changes are attributable.
- Give agents a one-click 'why this reply' view showing retrieved sources and tool results.
Evaluation and limitations
Evaluate the assist workflow, not only automation:
- Suggestion acceptance: share of drafts sent with minor or no edits.
- Handling time: median time to resolution for assisted versus unassisted tickets.
- QA score: human rubric scoring of accuracy, tone and policy compliance.
- Auto-send accuracy: audit sampled auto-sent replies; any hallucinated commitment is a release blocker.
- Escalation precision: whether the system handed off the right tickets.
Honest limits: helpdesk APIs differ in what they expose, moderation and audio endpoints are coming soon, PII handling is your responsibility, and auto-send multiplies the cost of every mistake — keep it narrow for a long time.
Honest comparison
| Capability | Support assist on Plugsky | Helpdesk AI add-on | Fully manual support |
|---|---|---|---|
| Integration | OpenAI-compatible calls in your service layer | Vendor plugin | Human workflow |
| Model choice | 30+ aliases, routed by intent | Vendor-selected model | Not applicable |
| Ticket state | Stays in your helpdesk | Vendor storage | Helpdesk |
| Escalation | Your rules and tools | Vendor-defined | Human judgment |
| Cost shape | Flat monthly self-serve, unlimited fair use | Per-seat add-on | Headcount |
Frequently asked questions
Should we auto-send AI replies?
Not initially. Start with suggested replies, measure acceptance and edit rates, and only auto-send narrow intents after sustained accuracy in human review.
Where should ticket state live?
In your helpdesk. The API call is stateless, so your service layer owns context assembly and writes results back to the ticket.
How do we handle PII?
Apply your policy before the call: redact or tokenize identifiers the model does not need, and keep logs aligned with retention rules for your regions.
Can we reuse existing OpenAI support scripts?
Yes. The endpoint is OpenAI-compatible, so changing the base URL and model names is enough for most integrations; verify streaming and tool-call paths afterwards.
Which model should draft replies?
plugsky-micro or plugsky-lite handles routine intents, while a stronger alias improves complex troubleshooting and multi-message context.
How do we measure value?
Suggestion acceptance, handling time for assisted tickets, QA rubric scores and escalation precision. Deflection is a bonus metric, not the goal.
Can we pilot on the free plan?
Yes. Two free models with no card cover a pilot, and the 14-day full-access trial lets you evaluate quality on real ticket samples.