AI agent platform
Research a topic, prepare documents and hand recurring tasks to AI agents. Set the goal and approve important actions. Start free — no credit card.
Tablet
Turns your ideas into something you can share.
Tell your agent what you need. Add your notes, files and any constraints.
Work in a conversation. Answer questions and connect tools only when the task needs them.
Check sources and the draft. Ask for changes before using or sharing the result.
Start with a small task and a result you can check. Here are three ways to put Snaga to work.
Compare options before making a decision. Name the products or topic and the criteria that matter to you. Ask for a comparison table with source links, then check the evidence and request any missing details.
Turn your notes into a document you can use. Provide the key points, audience and desired format. Ask for a draft and a Word document, then refine the wording in the same conversation before sharing it.
Keep up with a topic without repeating the same request. First check a sample summary. Then set the topic, sources and schedule for future updates, delivered through notifications. Review the results and adjust the task as your needs change.
Each plan includes a monthly allowance for agent work. Choose by your workload; extra usage, where available, requires a spending cap.
Try your first tasks. No card required.
No credit card · choose your first task
For individual work and regular tasks.
Cancel anytime · 14-day refund
For team work
For larger workloads and teams.
Cancel anytime · 14-day refund
For teams running fleets in regulated environments.
Billed monthly · Bring-your-own-cloud option
Tokens measure the text a model reads and writes, including the context it uses. Longer tasks and repeated steps use more. A token allowance does not guarantee a fixed number of finished tasks.
Where a daily limit is listed, it applies alongside the monthly allowance. Unused daily tokens do not carry over, so the full monthly amount cannot be spent in one day.
Limits apply together. A run is a task execution; its token usage depends on the model, the context and the work it does.
Scroll sideways to compare all plans.
| Included work | Free | Starter | Pro | Enterprise |
|---|---|---|---|---|
| Tokens per month | 2,700,000 | 22,500,000 | 100,000,000 | 2,000,000,000 |
| Included tokens per day | 90,000 | 750,000 | 3,333,333 | No separate daily limit |
| Runs per month | 300 | 3,000 | 15,000 | 200,000 |
| Runs at the same time | 2 | 4 | 10 | 50 |
| Agents | 2 | 3 | 20 | 200 |
| Teams | 1 | 1 | 5 | 40 |
| Images per month | 100 | 600 | 500 | 5,000 |
All plans include the full Snaga runtime · Limits, not feature gates · Cancel anytime
Meet a public guide on the real Snaga platform. Ask how to approach your task or which plan fits. Your own files and work belong in your private workspace.
Clear context, the tools you need and a result you can review.
Choose what an agent may access and which important actions need your approval.
Connect the services needed for your task. The connector catalog shows what is available on this platform.
Add notes, documents and reference materials so your agent has the information it needs.
Follow the work in your chat, ask questions and refine the result without starting over.
Review usage in your workspace. Plan allowances and extra usage rules are shown in the pricing section.
Use the hosted product, or deploy the open-source platform with your own provider configuration.
Three ways to put agents to work: buy a team that ships assembled, call one vendor's assistant API, or build it on a framework. The right answer depends on what you are willing to own.
| Capability | Snaga | Assistant API | Build it yourself |
|---|---|---|---|
| Every action that leaves the sandbox waits for a person, and the gate fails closedHow approvals work | yours to build | ||
| Multi-tenant API with roles, so a workspace is scoped from the first request | one tenant | yours to build | |
| Web, iOS and Android apps, and a browser — the same team on every surfaceThe app | |||
| MCP in both directions: your agents consume tools, and are reachable as oneWhat it connects to | consume only | consume only | |
| Spend caps per agent and per workspace, metered through Stripe | yours to build | ||
| Per-connector consent and the narrowest scope that works, with the live list published rather than assertedThe list | yours to build | ||
| Runs on your own machine, in your repo, with your toolchainSnaga Code | |||
| Self-host the whole platform inside your own perimeter |
A table where one column is all ticks is an advertisement, not a comparison. These are the reasons to not pick Snaga, written by the people who would rather you picked it.
Any model, any topology, any component swapped for your own. Snaga ships an opinion about how a team should work — the approval gate, the plan-build-verify loop, the sandbox — and you live inside it. If your design fights that opinion, the framework is the honest choice.
We hold no SOC 2 report and no ISO certificate today, and we say so on the security page rather than in a footnote. If procurement needs one on the desk this quarter, that is a real reason and not a negotiable one.
What we do and do not claimIf what you want is a single assistant inside a product you already run, an SDK is less to adopt and less to keep running. A platform earns its cost when there are several agents, several people approving their work, and a bill somebody has to explain.
Everything the product does, the API does. Five client libraries, generated from one OpenAPI document.
The same operations, models and event streams exist in every language — they are generated from the document the API itself is validated against, not written by hand alongside it.
uarp-sdknpmuarp-sdkcrates.ioSnaga-AI/uarp-swiftSwiftPMai.snaga:uarp-sdkMaven CentralIf an endpoint exists in one SDK it exists in all of them, because a generator put it there. No language is a second-class port.
Every mutating call carries an idempotency key, which is what makes it safe to retry. Transient failures back off with full jitter and honour Retry-After.
Eleven SSE endpoints resume with Last-Event-ID after a dropped connection, with a reconnect budget so a flapping server cannot spin the loop.
Failures arrive as RFC 9457 problem documents and decode into typed errors carrying the status, the detail and any field-level validation failures.
How to get started, what is free, and how your work stays under your control.