Key facts
| Agent runtime | Tool loop on /v1/chat/completions with streaming and function calling (live) |
| Model tiering | plugsky-micro and plugsky-lite for triage; plugsky-pro and plugsky-frontier for reasoning |
| Access control | Scoped API keys, RBAC and SSO/SCIM for workforce identity |
| Residency | Region-locked planes: EU (Frankfurt), GCC (UAE), APAC (Singapore), US (Virginia) |
| Audit | Inference metadata, key lifecycle and admin events exportable to SIEM |
| Human approval | Payment release, KYC outcomes and customer messaging stay behind your approval gates |
| Data handling | Configurable prompt retention; minimise PII before it reaches the model |
| Endpoint status | Agents on chat completions are live; assistants and responses endpoints are coming soon |
TL;DR
- Start with triage, summarisation and drafting; keep execution behind human approval.
- Route cheap classification to plugsky-micro and escalate hard cases to frontier models.
- Pin inference, embeddings and logs to one region-locked plane to satisfy residency.
- Scoped keys plus SIEM-exportable events make every agent action reviewable.
- Build on the live chat completions API; assistants-style endpoints are roadmap items.
How it works, step by step
- Pick one bounded workflow — dispute triage, KYC document review or fraud alert summarisation — and write down its allowed and forbidden actions.
- Create a Plugsky workspace and issue scoped API keys per environment, per team, per environment purpose.
- Implement the loop: send messages with a tools array, execute any tool_calls the model returns, append results and call again until it answers.
- Add approval gates so agents draft and recommend, while a named human releases payments or closes cases.
- Pin the workspace to a region-locked plane and set prompt retention to the minimum your policy allows.
- Stream inference, key and admin events to your SIEM, and run a regression suite against known historical cases.
- Pilot with a bounded user group, track false positives and escalate paths, then widen scope when evaluations hold.
Try it yourself
Open the agent workflow designer →
Where agents fit in a bank's workflow
The safest first agents are read-heavy and write-light. Fraud alert summarisation, dispute triage, KYC document extraction, policy search and analyst drafting all reduce handling time without giving the model authority. In Plugsky terms this is a loop around /v1/chat/completions: you pass a tools array, the model returns tool_calls, your code executes them against internal systems and appends the results until the model produces a final answer.
Because the API is OpenAI-compatible, the loop runs inside your existing service boundaries. The agent never talks to core banking directly; your orchestration service holds the credentials and enforces the permissions that already exist.
Controls, residency and auditability
Regulated deployments need three properties: identity, residency and evidence. Plugsky supports scoped API keys, RBAC and SSO/SCIM so agent services authenticate as themselves rather than as a human account. Inference, embeddings and logs can be pinned to region-locked planes — EU (Frankfurt), GCC (UAE), APAC (Singapore) and US (Virginia) — and where in-country processing is mandatory the same API runs in your VPC, on-prem or air-gapped.
- Audit: authentication, key lifecycle, admin changes and inference metadata are exportable to SIEM tools such as Splunk, Sentinel, QRadar and Datadog.
- Data minimisation: redact account numbers and identifiers before prompting, and set retention per workload.
- Keys: enterprise plans support customer-managed keys through KMS or HSM for a revocable kill switch.
Model routing, evaluation and rollout
Not every step deserves a frontier model. Use plugsky-micro or plugsky-lite for classification, routing and extraction, then escalate ambiguous or high-value cases to plugsky-pro or plugsky-frontier. This keeps latency and capacity predictable while preserving quality where it matters, and 30+ models sit behind one endpoint so switching tiers is a configuration change.
Before any widening of scope, build an evaluation set from historical cases with known outcomes. Measure precision on escalations, not just answer quality, and log every tool call so reviewers can reconstruct the decision. Start with analysts in the loop, then move to sampled review only when the evidence supports it.
Honest comparison
| Concern | Plugsky agents | Assistants-style platform | Building in-house |
|---|---|---|---|
| Build model | Tool loop on OpenAI-compatible chat completions | Managed threads and runs | Full platform build |
| Model choice | 30+ models, mix per step | Vendor catalogue only | You host and serve each model |
| Residency | Region-locked planes plus VPC, on-prem and air-gapped options | Usually a few global regions | Wherever you deploy |
| Audit | Identity, key and inference events exportable to SIEM | Vendor-controlled logs | Your own implementation |
| Human approval | Enforced in your application layer | Limited vendor approval features | You build every control |
| Status today | Live on chat completions | Varies by vendor | Months of engineering |
Frequently asked questions
Can Plugsky agents move money or execute transactions?
No. Agents can prepare, recommend and draft, but execution stays in your systems behind approval gates. Plugsky has no access to core banking and never executes payments.
Which workflows are realistic first targets?
Alert summarisation, dispute triage, document extraction, policy search and analyst drafting. These are read-heavy and write-light, so value appears before risk does.
How do we keep data in our jurisdiction?
Pin the workspace to a region-locked plane and use VPC, on-prem or air-gapped deployment where in-country processing is mandatory. Keys and logs should stay in the same jurisdiction as data.
What gets written to the audit log?
Authentication and key lifecycle events, administrative and policy changes, and inference metadata such as model, token counts and timestamps. Prompt retention is configurable per workload.
Does function calling work with streaming?
Yes. Streaming and OpenAI-style function calling are live, so you can show progress to analysts while tools execute server-side.
Do we have to use one model for the whole agent?
No. Route classification and extraction to plugsky-micro or plugsky-lite, and reasoning or synthesis to plugsky-pro or plugsky-frontier through the same endpoint.
How do we start without a procurement cycle?
Create a free workspace with plugsky-micro and plugsky-lite and no card, build the loop against your own evaluation set, then move to a paid plan or the 14-day full-access trial as evidence accumulates. See the live pricing page for plan details.