PromptShuttle
PromptShuttle is a server-side agent orchestration gateway for platform teams, agencies and software vendors. It is a drop-in OpenAI replacement: one API call spawns sub-agents, tool calls and multi-provider routing, so no agent SDK ever enters your code.
What is PromptShuttle?
PromptShuttle is an LLM orchestration platform that sits between an application and the major model providers. Its architectural stance is stated plainly: an API, not a framework. Where an agent SDK puts orchestration logic inside application code, so that changing an agent means redeploying, every team rebuilds the same plumbing and each one handles providers, retries and costs itself, PromptShuttle moves that logic into a proxy. Changing an agent becomes a configuration change, and a single deployment serves every team in the organization.
The entry point is an OpenAI-compatible chat completions endpoint, with the base URL app.promptshuttle.com/api/v1. Existing clients keep working: point them at that base URL with a PromptShuttle key. The twist is that the model name passed in the call can designate a flow rather than a model. Ask for research-agent and the platform spawns sub-agents, routes each one to a provider, runs the tool calls and returns a single aggregated answer. Flows are DAG workflows: a template can invoke other templates as sub-agents, with full tracing, cost tracking and a depth limit.
Prompts are parameterized templates written with [[parameter_name]] placeholders, detected automatically, whose versions are activated per environment (dev, staging, production). Tools come in four shapes, external HTTP tools, caller-hosted tools, MCP tools and agent tools, and the tool-calling loop is handled by the platform. MCP support runs both ways: PromptShuttle imports tools from external MCP servers, and exposes its own MCP server so the platform can be driven from Claude Desktop, Cursor, Windsurf or Claude Code through a single URL, with no installation and no CLI.
Around the runtime sits a dashboard: a per-request execution graph, flow analytics covering invocation trends, model breakdown and percentile latencies, runtime, cost and tool-call distributions, and costs aggregated by model and by flow. Multi-tenancy isolates keys, billing and flow configuration per team or client. Seven providers are documented, namely OpenAI, Anthropic, Google, Groq, DeepSeek, xAI and Perplexity, behind a public catalog of 44 models listing context window, maximum output and price per million tokens. Four named customers appear on the site: TenderStrike, MentorMode, BrandKarma and Optivise. What the platform does not ship is ready-made agents: the flows and prompts are yours to write.
What it does
- Orchestrate multi-agent workflows server-side from a single API call
- Replace OpenAI with a compatible endpoint by changing only the base URL and the key
- Route each agent to the best model across seven providers without touching application code
- Attach tools to agents: HTTP webhooks onto existing ERP or CRM software, or tools imported from external MCP servers
- Version prompt templates and promote them across dev, staging and production environments
- Cap spending per request, set per-tenant budgets and trigger cost alert webhooks
- Isolate teams or clients with separate keys, billing and flows, and trace each run through its agent tree
When to use PromptShuttle / When not to
A quick filter to help you decide if PromptShuttle is the right fit.
When to use PromptShuttle
- Platform teams opening LLM access to several internal teams, with per-team API keys, spending budgets, guardrails and an audit record for every request
- Agencies and consultancies running AI for multiple clients, using the built-in multi-tenant isolation of keys, billing and flow configuration
- Vendors of existing software (ERP, CRM, internal tools) that must become agent-capable through webhooks rather than through a rewrite
- Engineering teams already on the OpenAI SDK who want to change model or agent without redeploying: only the base URL and the key change
- Background jobs and schedulers such as Trigger.dev or Temporal that call AI as a plain HTTP service, with no SDK to install
When not to use PromptShuttle
- Non-technical users: everything runs through an API key and an HTTP endpoint, and there is no consumer-facing application
- Anyone looking for ready-made agents: no agent ships with the product, you define your own flows, prompts and tools
- Mobile, desktop or browser-extension users: the product exists only as a web dashboard and an API
- Buyers bound by a compliance review: no SOC 2 or ISO 27001 certification, no declared hosting country and no data processing agreement are published
- Teams that need a fully identified vendor or long log retention: no postal address or legal entity form is published, and logs are kept 30 days on Team and 3 months on Business
How to use PromptShuttle
A typical end-to-end flow, from setup to results.
- Create an account on app.promptshuttle.com and generate an API key
- Fastest route, announced at under two minutes: point an existing OpenAI client at the base URL https://app.promptshuttle.com/api/v1 using that key, with curl, Python and TypeScript samples provided
- If you have no SDK at all, a plain HTTP POST is enough; the legacy software page adds curl, Python, Java and C# examples
- For an orchestrated flow, open Flows > Create Flow in the dashboard, where a slug is generated automatically
- Write the prompt template, marking parameters as [[parameter_name]]; they are detected automatically
- Select a model, for example openai/gpt-4o, and save the version
- Activate that version for an environment (dev, staging or production) in the flow settings
- Call the flow exactly as if it were a model name in a standard chat completions request
- Register an external MCP server with POST /api/v1/mcp-servers (name, url, optional headers), then discover, import and assign its tools
- Set the guardrails: maxRequestCostCredits, maxAgentDepth (system default 10), costAlertThresholdCredits and alertWebhookUrl; optionally add https://app.promptshuttle.com/mcp to an MCP client to drive the platform from your editor
Pros & Cons
Pros
- Migration is close to trivial for anyone already using the OpenAI SDK: a base URL and a key are the only changes
- Agent logic leaves application code, so changing an agent is a configuration change rather than a redeployment, and one deployment serves every team
- Unusually fine-grained cost control: per-request caps, per-tenant budgets, webhook alerts and the cost of the call returned in the response
- Seven providers behind a single integration, with aliases and fallbacks, plus a public model and price catalog that is also readable through the API (GET /api/v1/models/descriptors)
- Native observability of the agent tree, with no instrumentation to write
- Multi-tenancy and isolation available from the start, which is exactly what agencies and platform teams need
- No SDK required, so a plain HTTP POST works from Java, C#, an ERP or a job runner, and the platform itself can be driven in natural language from Claude Desktop, Cursor, Windsurf or Claude Code through its MCP server
Cons
- Corporate transparency is minimal: no postal address and no legal form, the contracting party in the Terms is simply "PromptShuttle", there is no about page (docs.promptshuttle.com/about returns 404) and no team is presented
- The terms of service and privacy policy sit on a single page with no effective date and no version number
- No data processing agreement is offered, no sub-processor list is published, no Article 27 EU representative and no DPO are named, no hosting country or region is declared, and no security certification is displayed (neither SOC 2 nor ISO 27001)
- Nothing is said about whether customer data is used to train models, and no opt-out is documented
- The license granted over user content is very broad, covering modification, derivative works, distribution and sub-licensing
- A token surcharge stacks on top of the subscription, 15% on Team and 5% on Business, log retention is short (30 days on Team, 3 months on Business) and the Team tier is capped at 3 users
- Prices are rendered by a JavaScript Stripe table and never appear in the page HTML, the only contact channel is a single obfuscated email address, and the governing law is that of the State of New Mexico, with no alternative forum for European customers
Pricing & Plans
No permanent free plan is described. The homepage advertises "Start free" and a "Get Started Free" button, yet the Stripe table that feeds the pricing section carries only two paid tiers, and no free trial is configured on any of its four prices. The lowest entry point is therefore Team at USD 29.00 per month, or USD 149.00 per year; Business is priced at USD 100.00 per month or USD 1,000.00 per year. The subscription does not cover consumption: usage is billed in credits bought through Stripe at 1,000,000 credits for USD 1.00, credited immediately. The cost of a call is input tokens multiplied by the input rate plus output tokens multiplied by the output rate, with reasoning tokens, cached tokens, cache writes and tool costs added depending on the model; published model rates range from USD 0.20 to USD 30.00 per million input tokens. A surcharge then applies on top of token cost: 15% on Team and 5% on Business.
- up to 3 users
- unlimited requests
- 30-day log retention
- 15% surcharge on tokens
- presented as being for smaller teams or call volumes
- unlimited users
- unlimited requests
- 3-month log retention
- 5% surcharge on tokens
- presented as being for larger organizations or call volumes and flagged as the best deal
- Consumption credits
- bought separately in packs through Stripe at 1
- 000
- 000 credits for USD 1.00
- No free tier and no quote-based enterprise tier appear in the pricing table
Data, GDPR & hosting
A consolidated view of how PromptShuttle handles your data.
GDPR overview
The Legal page carries an explicit GDPR section for EU and EEA residents. It sets out rights of access, rectification, erasure, objection to processing, restriction and portability, states that consent can be withdrawn at any time where processing relies on it, and points users to a data protection authority for complaints. A separate CCPA section covers California, with the right to know, the right to deletion subject to exceptions and non-discrimination. Both sections commit to resolving complaints and to not selling personal data. The gaps are documentary: no EU representative is designated under Article 27, no DPO is named, no data processing agreement is offered, no sub-processor list is published, and no hosting country or region is declared. The privacy contact is the generic support@promptshuttle.com address, obfuscated on the page, and the policy carries no effective date and no version number.
Who owns the data?
Under the Terms, users keep ownership of the intellectual content they create with PromptShuttle. In exchange, using the Service grants PromptShuttle a non-exclusive, transferable, sub-licensable, royalty-free and global license to use, store, display, reproduce, modify, create derivative works from and distribute that content. The limit set on this license is one of purpose rather than scope: it applies solely to operating and improving the Service. Because the license is transferable and sub-licensable, other parties can be brought under it, and prompts necessarily reach the model providers behind the platform. Separately, the publisher states that it does not sell or lease personal information to third parties.
Reuse rights
Users retain ownership of the content they create, so outputs are theirs to reuse without asking permission; the license runs the other way, from the user to PromptShuttle, and is limited to operating and improving the Service. On the collection side, signing up requires a name, an email address and a password. The platform also records usage data: the applications created, the tools used and interactions with the platform, along with cookies and tracking technologies for user experience and traffic analysis. The declared purposes are providing, maintaining and improving the service, analyzing usage trends, sending updates and security alerts, and communicating about products, offers, promotions and events. Data tied to the applications created is stored and analyzed, and the policy states that "the analysis is strictly for service enhancement and will not be used for any independent purposes". Sharing with third-party service providers is possible under strict confidentiality agreements, though no sub-processor list is published. Prompts also transit through the model providers behind the platform (OpenAI, Anthropic, Google), sometimes hosted in turn by Groq or Replicate. Nothing in the documents addresses model training on customer data, in either direction.
Data retention & training
Hosting summary
No hosting country and no hosting region is declared anywhere: not in the terms of service, not in the privacy policy, not in the documentation. There is no data residency commitment and no hosting certification. What can be observed is only network-level. The domain is served behind Cloudflare (AS13335) on the anycast address 104.21.0.231, which identifies a CDN edge node rather than a place of storage, so nothing about where data actually sits can be inferred from it. The API and the dashboard are served from app.promptshuttle.com. Beyond the platform's own infrastructure, the LLM processing itself is delegated to OpenAI, Anthropic and Google, sometimes hosted in turn by Groq or Replicate, so prompts and responses transit through those providers' infrastructures. The registrar record places the domain with Porkbun LLC in the United States and the Terms designate the law of the State of New Mexico, but neither of those is a statement about where data is processed or stored. Any organization with a residency requirement will have to obtain this information from the publisher directly.
Things to keep in mind
Risks and trade-offs to weigh before adopting PromptShuttle.
- The subscription is not the bill: consumption is billed in credits on top of it, with a surcharge on tokens of 15% on Team and 5% on Business, which has to be built into any cost estimate
- An unattended agent tree can multiply calls quickly, so the per-request cost ceiling (COST_LIMIT_EXCEEDED) and the agent depth limit (default 10, DEPTH_LIMIT_EXCEEDED) need to be set deliberately rather than left at their defaults
- Log retention is short and tier-bound, 30 days on Team and 3 months on Business, which limits incident investigation after the fact; the Team tier is also capped at 3 users
- The license granted over user content allows modification, derivative works, sub-licensing and distribution, and prompts often carry commercially sensitive material
- No data processing agreement, no sub-processor list and no declared hosting country make a compliance-bound purchase difficult to document
- Prompts and data transit through third-party providers, themselves sometimes hosted by yet other parties such as Groq or Replicate, which widens the chain of custody beyond the publisher
- The legal documents carry no effective date, so changes leave no visible trace, displayed prices live in a JavaScript Stripe table absent from the page HTML, the publisher has no published legal form or address, and disputes fall under the law of the State of New Mexico
Setup & Integrations
Technical difficulty
Developer-level, but the floor is low. An API key is required. The shortest path is advertised at under two minutes: change the base URL and the key of an existing OpenAI client. No SDK is needed: a plain HTTP POST works, with samples in curl, Python, TypeScript, Java and C#. Building a flow is dashboard work: title, template with [[param]] placeholders, model choice, activation per environment. Connecting existing software is more demanding: HTTP endpoints must be exposed and registered as tools. Guardrails are set per tenant or per request, and an MCP client can drive everything from the editor.
Deployment
Integrations
Behind PromptShuttle
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Do I have to rewrite my application to use PromptShuttle?
Do I need to install an SDK?
Which model providers can I reach through it?
How much does PromptShuttle cost?
Is there a free plan or a free trial?
How do I stop an agent loop from running away?
Can I connect an existing ERP or CRM to it?
Does it work with MCP servers?
How long are logs kept?
Is my data used to train models?
Should you pick PromptShuttle?
PromptShuttle is infrastructure for engineering teams, not a product for end users. Its thesis is consistent from end to end: agent logic does not belong in application code, it belongs in a configurable proxy. Everything else follows. The OpenAI-compatible endpoint makes migration almost free for anyone already using the OpenAI SDK. Cost control is unusually granular for this category, with per-request caps, per-tenant budgets, alert webhooks and the cost returned in each response. Observability of the agent tree comes without instrumentation to write, and multi-tenancy is there from the start rather than bolted on, which is precisely what agencies and platform teams need.
The weakness sits on the other side of the ledger, and it is documentary rather than technical. The publisher is identified only as PromptShuttle, with no legal form, no postal address, no about page and no team presented. There is no data processing agreement, no sub-processor list, no declared hosting country, no security certification and no Article 27 representative, and the legal documents carry no effective date. Nothing is said about whether customer data trains models. For a team that has to clear a compliance review, that silence is the obstacle, not the technology.
Pricing is clear once found, but it is rendered in JavaScript rather than in the page HTML, and the subscription is only part of the bill: consumption is billed in credits with a surcharge on tokens of 15% on Team and 5% on Business.
The fit is narrow and real. A platform team opening AI to several internal teams, an agency serving multiple clients, or a vendor making existing software agent-capable without rewriting it will find here something that few competitors offer as a service rather than as a library.
- Choosing a selection results in a full page refresh.
- Opens in a new window.