What’s next
Directional, not a commitment — the order changes more often than the list does.Soon
- SSO and OIDC — including Microsoft Entra sign-in.
- Audit logs — who changed what, and when.
- LangChain integration — a package for using PromptJuggler prompts and workflows from LangChain.
- Agent skills — reusable skills your agents can draw on.
- Agent memory — a document store agents can write to and read back, so what they work out in one run is still there in the next.
- Evaluation improvements — datasets built from production runs rather than hand-written cases, JSONL import, assertions on metrics like cost and duration, and evals surfaced on the prompt and workflow pages they belong to, with better filtering and sorting.
Later
- Dashboard — usage and spend, recent runs, and how each prompt and workflow is actually behaving, gathered into one place.
- Batch and priority processing — provider batch queues and flex/priority tiers, across OpenAI, Anthropic and Gemini.
- Provider-side code execution — let the model run code in the provider’s own sandbox, alongside the script tool you already have.
- Sub-tenants — workspaces nested under a parent, for agencies and resellers.
- Automatic conversation summarisation — keep long threads inside the context window without losing the plot.
Missing something?
Customer demand is the main thing that moves items up this list. If something here matters to you, or something that isn’t here matters more, tell us.What’s shipped
- Token streaming — tokens reach the browser as they’re generated, over SSE, with replay if a connection drops. Opt-in per prompt, so nothing changes unless you ask for it.
- The browser SDK —
@promptjuggler/browser, a streaming client for the people using your product rather than for your backend. A React hook and an Angular signals adapter sit on a framework-agnostic core. - The emit tool — emit lets a model hand your application structured, schema-validated payloads mid-conversation, instead of leaving you to parse them back out of prose.
- Five backend SDKs — Python, TypeScript, Java, PHP and Go, generated from the OpenAPI spec and then wrapped in hand-written facades so they don’t feel machine-made.
- Trash and recovery — deleting a prompt, workflow or knowledge base moves it to the trash rather than destroying it. You can bring it back.
- Per-environment provider keys — point staging and production at different OpenAI, Anthropic or Gemini credentials, and each run picks up the right one.
- Knowledge bases and RAG — upload documents and PromptJuggler handles extraction, chunking, embedding and semantic search, as both a tool your prompts can call and a workflow node. No vector database to run.
- Sub-agents — prompts and workflows became callable as tools from other prompts. An agent can hand work to another agent, which can hand it to another, nested as deep as the problem needs.
- Threads and forking — threads got their own view, including forking a conversation from any point.
- Loop, array and conditional nodes — loop, array-operation and conditional nodes, for the logic that shouldn’t need a model.
- Environment variables — configuration scoped per run, so one prompt can point at different systems in staging and production.
- Anthropic models — Claude joined OpenAI and Gemini, with extended thinking supported.
- MCP tools — any MCP server can be connected as a tool, across every provider.
- Redesigned interface — the UI was rebuilt on a lighter component set, with clearer navigation and a sidebar you can actually find things in.
- JMESPath — the JSON Query node accepts JMESPath alongside JSONPath, so you can use whichever syntax you already know.
- Versioning rebuilt — the draft-and-publish flow: work in a draft, publish when it’s ready, and leave your callers undisturbed until you do.
- Published OpenAPI spec — the full API surface, specified. Later the foundation every SDK is generated from.
- JSON schema editor — define structured input and output shapes without hand-writing schema.
- GitHub sign-in — one fewer password.
- Conversation history — prompts stopped being one-shot and started keeping the thread.
- Pin to a tag — call a specific published version instead of whatever happens to be newest.
- Evaluations — run a prompt against a dataset, assert on the results, and score them. When you change a prompt you find out whether it got better or worse, rather than squinting at a handful of outputs and hoping.
- Cost tracking — every run reports what it cost.
- Reasoning controls — reasoning effort, reasoning summaries and verbosity, exposed and normalised across providers.
- Google Gemini — a second provider alongside OpenAI.
- File uploads — attach files to a run.
- Script execution — run sandboxed code as a step, for the glue work that doesn’t need a model.
- In production — PromptJuggler started serving production workloads: prompts with versioning and OpenAI model selection, visual workflows with fan-out and fan-in, tools, webhooks, and a public REST API.