Multi-model AI workspace
Your AI workspace.
From questions to completed tasks.
Chat with connected models, explore your documents, and run tool-assisted workflows in one workspace — with citations you can open and approval before anything leaves it.
simulated modelNo provider key is configured, so replies come from a deterministic built-in model.
No sign-up needed — or sign in.
How it works
Four steps, start to finish
The workspace is built around one loop: ask, choose, watch, verify.
- 01
Ask a question, or add a document
Type a question in Chat, or upload a PDF, text or Markdown file. Indexing runs in the background and the row shows its real status.
- 02
Choose a model and the tools it may use
Pick from the models the server actually reports. Read-only tools run freely; anything that writes outside the workspace is gated.
- 03
Follow progress, and decide when asked
Each step is recorded as it happens — plans, tool calls, results. If a consequential action comes up, the run pauses for you.
- 04
Get an answer with its sources
Answers over your documents carry citations that open the exact passage used. If the evidence is not there, it says so instead of guessing.
Features
What is actually in the workspace
Each capability below is implemented in this build. Where something needs configuration before it can do real work, it says so.
Documents with citations
Answers that show their evidence
Ask a question across your own files. Retrieval finds the passages, and the citation on the answer opens the exact chunk it came from — not a page reference, the text itself.
- PDF, plain text and Markdown, chunked and embedded on upload.
- Hybrid keyword + vector retrieval over your documents only.
- Refuses to answer when the evidence is not there, rather than guessing.
Tool-assisted tasks
A run you can actually follow
Give the agent a goal. It plans, calls tools, and records every step as it happens. Nothing in the timeline is decorative — it is the run's real event log.
- Step, tool call and tool result events, persisted per run.
- Step, time, token and cost budgets end a run with a stated reason.
- Cancel a run mid-flight; partial work is kept.
- Planbreak the goal into steps
- Tool calldocument_search
- Tool result3 passages matched
- Approvalpaused for your decision
- Answerwritten with citations
Approval controls
Consequential actions wait for you
Read-only tools run freely. Anything that would write outside the workspace pauses and shows you the destination and effect before it happens.
- Approval requests record the tool, destination and effect.
- Idempotency keys stop a duplicate send on resume.
- Every decision is kept in the run's history.
- tool
- send_external_message
- effect
- Sends a message outside the workspace
Multi-model conversations
Pick the model per conversation
The selector lists the models this server reports — nothing more. Add a provider key and real models appear alongside the built-in deterministic one.
- Capability registry: streaming, tools, vision, context window.
- Keys stay server-side and never reach the browser bundle.
- Outbound hosts are allowlisted to prevent SSRF.
User-controlled memory
Short preferences you write yourself — tone, format, project context. Nothing is remembered unless you put it there, and every entry can be disabled or deleted.
Edit demo memoryliveMCP connections
Attach Model Context Protocol servers to give the agent extra tools. Tool definitions are hashed and require approval, so a server cannot silently change what it can do.
See connectionsneeds a server configuredProduct preview
Look around before you sign in
These panes show the sample content seeded into every demo session, so the workspace you open will look like this.
What are the success metrics in the Orchard redesign brief?
The brief sets four success metrics, measured ninety days after general release:
- Basket abandonment below 50%, down from 68%.
- Median checkout completion under 90 seconds.
- Shop-follow actions per active user up at least 25%.
- Crash-free session rate at or above 99.5%.
Uploads are capped in the demo; your own files stay in your own session.
Entries are yours to change - add, edit, disable or delete any of them.
Use cases
Four things to try first
Each of these works in the demo with the content already seeded.
Summarise a project document
Summarise the Orchard redesign brief in five bullets.
A short summary drawn from the indexed brief, each claim traceable to a passage.
Answer with source citations
What are the success metrics, and where are they stated?
An answer with [1]-style citations that open the exact chunk used.
Find information with a tool
Search my documents for the delivery risks.
A document_search tool call, its result count, then an answer built on it.
Save a preference
Always answer in British English, conclusion first.
A memory entry you can edit or delete, applied to later conversations.
Control and privacy
What the implementation actually guarantees
Only behaviour that is implemented and tested in this build. No compliance claims, no certifications.
- Workspace data isolation
- Every query is scoped to the signed-in account. Ownership comes from the verified token, never from a client-supplied id, and a request for someone else's record returns 404. Integration tests cover conversations, documents, chunks, runs and memory.
- User-controlled memory
- Memory holds only what you write into it. Entries are listed in full, can be disabled or deleted, and are never inferred from your conversations behind your back.
- Review of consequential actions
- Tools are classified by consequence. A tool that writes outside the workspace cannot run until you approve that specific call, and the request shows the destination and effect.
- Server-side credential handling
- Provider keys live in server configuration and are never included in the browser bundle. Outbound model endpoints are checked against an allowlist to prevent SSRF.
FAQ
Questions, answered honestly
What is JONY-AI?
A multi-model AI workspace: conversations with a chosen model, question answering over your own documents with citations, tool-assisted tasks with an approval step, MCP integrations, and memory you control. It is a working application, not a mockup.
What can I try in the demo?
Everything in the workspace: chat, upload and query documents, start and cancel a task, approve or reject a gated action, and edit memory entries. The demo starts with two sample documents, a seeded conversation, four memory entries and one completed task already in it.
Which model providers are configured?
None on this deployment. The only available model is a deterministic built-in one, labelled simulated wherever it appears. Add an OpenAI, Anthropic or Gemini key in server configuration and real models appear in the selector automatically.
Can I upload documents?
Yes — PDF, plain text and Markdown. Files are chunked and embedded on upload, and indexing status is shown per document. Demo sessions have a smaller per-file size cap and a limit on how many documents they can hold.
What are MCP connections?
Model Context Protocol servers that give the agent extra tools. Tool definitions are hashed and must be approved, so a server cannot silently change what it is able to do. Managing connections is disabled in the public demo; no MCP server is configured on this deployment.
Which actions require approval?
Any tool classified as writing outside the workspace. In this build that is send_external_message, which is itself simulated and records to an internal outbox rather than contacting anyone. Read-only tools such as document search run without interruption.
What happens to demo data?
Each visitor gets a separate demo account, so your session's conversations, documents and memory are not visible to anyone else. Demo content is sample data kept apart from real accounts. Treat a demo session as disposable and do not upload anything sensitive to it.
Open a workspace and try it
The demo starts in a private session with sample documents, a conversation and a completed task already in place. No sign-up.