Key facts
| Model access | 30+ models behind one API key, from free tiers to frontier |
| Tool calling | Function calling, streaming and JSON mode are live |
| Retrieval | Embeddings and RAG are live for memory and grounding |
| Security | Scoped keys, RBAC, SSO/SCIM and audit logs are live |
| Deployment | Cloud, VPC, on-prem and air-gapped with region choice |
| Pricing | Flat monthly self-serve plans; see the live pricing page |
| Free tier | plugsky-micro and plugsky-lite free, no card required |
| Roadmap | Assistants, files, batch and moderation endpoints are coming soon |
TL;DR
- Start with model breadth and tool-calling reliability; everything else is secondary.
- Decide early between a managed runtime and a framework you host.
- Check security controls and deployment options before the pilot, not after.
- Prefer platforms that do not lock your orchestration into proprietary formats.
- Pilot on real tasks with a fixed evaluation set and a cost budget.
How it works, step by step
- Write down your top three agent use cases and the tools each one needs.
- Shortlist platforms that support function calling and the models you want.
- Check governance: scoped keys, RBAC, audit logs and data residency.
- Confirm deployment fit: shared cloud, VPC, on-prem or air-gapped.
- Run a two-week pilot on real tasks with a frozen evaluation set and cost tracking.
- Compare results on quality, cost per task and operational effort.
- Choose the option with the lowest exit cost if requirements change.
Try it yourself
Open the AI agent comparison →
The criteria that actually matter
Platform comparisons usually lead with model names, which is the least durable dimension. Models change quarterly. What persists is the integration surface: is the API OpenAI-compatible, does function calling work reliably across models, is there a real embeddings path, and can you move orchestration without rewriting it.
- Model breadth: enough tiers to route cheap work and hard work on one key.
- Tool calling: structured outputs, parallel calls and predictable schema handling.
- Retrieval: embeddings and RAG without a second vendor for the basic path.
- Observability: request-level logs, usage analytics and traceability per run.
- Governance: scoped keys, RBAC, audit logs and residency.
- Exit cost: standard protocols and data you can export.
Where platforms differ
Framework-first tools give you control and require you to operate everything. Managed agent products give you speed and constrain how agents are built. API platforms sit in the middle: they serve models and primitives, leaving orchestration in your code. Most teams settle on the middle because agent logic is product logic and should not live inside a vendor runtime.
Deployment is the great divider. Some platforms only run in the vendor's cloud; others offer region choice, VPC, on-prem or air-gapped operation. If data residency, procurement or export controls apply to you, this single criterion can decide the shortlist before quality comparisons begin.
A shortlist process you can run in a week
Pick three platforms that pass the governance and deployment filters. Build the same small agent on each: two tools, one retrieval source, a turn cap and trace logging. Run an identical task set and record success, steps, latency and cost per task. Two weeks is usually enough to see which platform gets out of your way.
Plugsky is a strong default for the runtime layer: 30+ models on one OpenAI-compatible API, live function calling, streaming, JSON mode, embeddings, RAG and agents, with scoped keys, RBAC, SSO/SCIM and audit logs, plus deployment from shared cloud to VPC, on-prem and air-gapped. Self-serve plans are flat monthly, and the free tier covers plugsky-micro and plugsky-lite. Current plans are on the live pricing page; assistants, files, batch and fine-tuning endpoints are coming soon.
Honest comparison
| Criterion | What good looks like | Red flag | How to test |
|---|---|---|---|
| Model breadth | Free and frontier tiers on one key | One model family only | Run the same task on three models |
| Tool calling | Parallel calls and schema fidelity | Frequent malformed arguments | Trajectory assertions on tool calls |
| Governance | Scoped keys, RBAC, audit logs | One shared key for everything | Attempt role separation in the pilot |
| Deployment | Cloud, VPC, on-prem option | Vendor cloud only | Ask for the deployment matrix |
| Exit cost | OpenAI-compatible and exportable | Proprietary agent format | Swap base URL in the prototype |
Frequently asked questions
Which platform is cheapest?
Cost depends on your task mix, not list prices. Measure cost per completed task in a pilot, then compare with the live pricing pages of the shortlisted options.
Do I need a managed agent platform?
Only if you want the vendor to own orchestration and memory. Teams with product-specific agent logic usually prefer an API platform and keep the loop in their own code.
How important is OpenAI compatibility?
It is mainly about exit cost. Compatibility means your SDK, tooling and framework integrations keep working if you change providers or run multiple.
What should I pilot first?
One high-volume, low-risk workflow with two or three tools and a clear success metric. That exposes integration friction quickly without risking production.
How do I judge observability?
Ask whether you can trace one run end to end: model, turns, tool calls, arguments, latency and cost. If the answer is a dashboard of aggregate counts only, keep looking.
Is data residency a blocker?
It can be. Plugsky offers region choice plus VPC, on-prem and air-gapped deployment, which covers most sovereign and regulated requirements.
Can I use more than one platform?
Yes, and it is a reasonable risk strategy. OpenAI-compatible interfaces make multi-provider routing practical without maintaining separate integrations.