Register your data model, metric definitions and business vocabulary once. Upsolve grounds every agent answer in them, versions them like code, and keeps column values fresh as your data changes.
\\n \\nSchemas don\\'t say which table is the source of truth or how your team defines a metric.
Prompt-only context drifts the moment a status, term or category changes in the data.
When the agent picks the wrong definition, the number looks right and nobody notices until the board meeting.
Attach tables and fields to an Upsolve data model, with descriptions, types, and primary and foreign keys.
Put canonical definitions, analysis approaches and output formats in a versioned system prompt.
Mark columns as selectable and Upsolve pre-caches their values, refreshed nightly or on your schedule.
Ten definitions of revenue live in docs, Slack and people\\'s heads. The agent picks one and the number looks right.
Tables, keys, descriptions and canonical definitions live in one versioned model the agent always reads.
Every table and column carries the context your best analyst would give a new hire.
System prompts and models have versions and drafts, so you can change context safely and roll back.
The agent already knows your five contract statuses and four payment terms before it writes SQL.
Start from your warehouse as it is today. Add definitions as you find the gaps.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nNo. You can build the data model in Upsolve directly on top of your warehouse tables, then add definitions over time.
Prompts hold business practice. The data model holds the truth of the underlying data: tables, keys, descriptions and live column values. Upsolve uses both.
Nightly by default, or on whatever cadence you set, so the agent\\'s picture of your data doesn\\'t drift.
Can\\'t find your answer? Get in touch.
\\nRow-level security enforced server side, before any query runs. Set rules from user and org properties, test them as any user, and deploy one agent to every tenant. The agent never sees the rules, so it can\\'t bypass them.
\\n \\nAgents that enforce their own permissions can be confused, prompt-injected or simply wrong.
Multi-tenant SaaS can\\'t put an agent in front of customers without hard data isolation.
Security reviews stall AI projects when nobody can show exactly what each user can see.
Give organisations and users properties such as country, tenant or role.
Global rules like column country = user.country apply to every table that has the column. Custom SQL rules cover the rest.
Preview any query as any user and see the exact WHERE clause Upsolve will apply before it runs.
The agent is told in its prompt not to show other tenants\\' data, and you hope it listens.
Every query is pre-filtered on the server from user properties before it runs. The agent never sees the rules.
Every query, whether an agent or a person wrote it, is filtered before it reaches your database.
Rules live outside the agent\\'s context, keeping deterministic security separate from agentic workflows.
Choose whether a missing property skips the rule or blocks the data entirely.
One agent, deployed to thousands of customers, each seeing only their own rows.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nNo. Rules are applied server side as pre-filters on every query, and the agent never sees them.
Write a custom SQL rule for that table. Upsolve reads the user property and applies your SQL before anything runs.
Yes. Live views on shared canvases re-run under the viewer\\'s own permissions.
Can\\'t find your answer? Get in touch.
\\nAttach any MCP server to your Upsolve agent. It still answers from governed warehouse data, and can now read and write where your team works: save an analysis to Notion, open a ticket, pull context from docs.
\\n \\nAnalyses get copied by hand into docs and tickets, and lose their context on the way.
Useful context lives in tools your data agent can\\'t reach.
Every new integration used to mean custom engineering work.
Paste the MCP server URL in project settings and add a one-line instruction, like \"use this to save analyses in Notion docs\".
Connect through OAuth, then allow all tools or block the ones you don\\'t want the agent to use.
Ask a data question, then tell the agent to write the analysis to Notion. It finds the right page and creates it.
A great answer stays in the chat window, and someone copies it into a doc by hand.
The agent answers from governed data, then writes the analysis to Notion or opens the ticket itself.
Notion, Linear, GitHub, PostHog and any other tool that exposes an MCP server.
Allow or block individual tools on every server you connect.
Tell the agent when and why to use each connection.
Questions still run through your data model and row-level security.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nFor actions and extra context, yes. For numbers across several sources, use the Upsolve Data Plane, which queries them with real SQL instead of combining results in the agent\\'s context.
Any server reachable by URL. Standard OAuth flows are supported for authentication.
Yes. After connecting, choose which of the server\\'s tools are allowed.
Can\\'t find your answer? Get in touch.
\\nWe rebuilt the Upsolve agent harness from scratch for people who don\\'t know the underlying data. Twice as fast, self-correcting, extendable with your own skills, and personalised for every user.
\\n \\nHardcoded agent workflows are slow and break when the data doesn\\'t look the way they expected.
Generic agents can\\'t learn the specific analyses your business runs.
Every user gets the same one-size-fits-all output.
The harness plans with tools and skills, writes SQL, validates it and renders the chart.
When the data surprises it, the agent recovers and keeps going instead of getting stuck.
Suggested skills and per-user memory improve every conversation that follows.
Hardcoded workflows stall when the data surprises them, and every user gets the same generic output.
A tool-and-skill harness that recovers from bad states, runs your analyses and remembers each user.
Tool calls and skills replace pre-hardcoded workflows.
Encode your own analyses, like a spend driver skill that frames the comparison and breaks down the change.
The agent proposes new skills from real conversations for you to approve.
Chart types, colours, time bucketing and more, remembered per user without changing accuracy.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nGenerate SQL and generate chart skills ship by default and improve over time. You add custom skills for your own analyses.
No. Suggestions appear for a builder to review and add to the agent\\'s skills.
Embedded in your product, in Slack, in Claude through MCP, or inside your own agent harness.
Can\\'t find your answer? Get in touch.
\\nFavourite any chart your agent makes, drop it on a drag-and-drop canvas and share it with your org. Teammates can fork the conversation behind any chart and keep digging, with live views that respect each viewer\\'s permissions.
\\n \\nChat answers get buried in history and can\\'t be revisited or reused.
Screenshots of charts lose the question and the context behind them.
Teams can\\'t build on each other\\'s analysis without starting over.
Save any chart. It keeps the full conversation, so you can reopen it and continue.
Drag favourites onto a canvas, resize and position them, and save versions.
Share with your organisation or publish a read-only snapshot. Teammates fork the chat behind any chart.
Great charts get buried in chat history, and screenshots lose the question behind them.
Charts are saved with their conversation, arranged on canvases, shared, and forked by teammates.
Every saved chart links back to the conversation that made it.
Build a board from charts created over weeks of analysis.
Pick up a teammate\\'s analysis and drill further in your own thread.
Snapshots show what the creator saw. Live views re-run under each viewer\\'s permissions.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nYes. Canvases work inside the embedded iframe, so your end users can save and share their own dashboards.
A read-only snapshot of the data at the time it was created. It never queries your database live.
Anyone who turns on the live view, filtered to the data their own permissions allow.
Can\\'t find your answer? Get in touch.
\\nEvery conversation, trace, SQL query and skill call your agent makes, in one place. Live LLM-as-judge evals flag wrong answers, lazy responses, empty results and frustrated users, so you can fix the context before trust breaks.
\\n \\nA slightly wrong number looks exactly like a right one.
Users who get a bad answer ask again, then stop using the agent.
Without traces, builders can\\'t tell whether the problem is the data, the context or the question.
The agent chat history shows every user, question, eval result and cost.
Open any session to see RAG retrieval, the skill used, the SQL, validation and the chart.
When an eval flags a problem, fix the prompt, skill or data model and watch the next conversations improve.
Wrong numbers and frustrated users go unnoticed until someone stops trusting the agent.
Every session is traced and graded on real usage, so you see the failure and the fix in one place.
Answer correctness, agent laziness, null or empty results, chart failures and user frustration.
The exact SQL, RAG sources and tool calls behind every answer.
See spend, charts generated and errors for each individual trace.
Embedded, Slack and Claude conversations all land in the same history.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nSignals like a user repeating the same request because they didn\\'t get what they expected.
On real conversations as they happen, so you see how the agent performs for actual users.
Yes. Every deployment surface writes to the same agent chat history.
Can\\'t find your answer? Get in touch.
\\nConnect hundreds of sources into one Upsolve data model, no warehouse required. Your agent writes real SQL across Stripe, PostHog, HubSpot, GitHub and your warehouse, with row-level security, instead of stitching MCP calls together.
\\n \\nWith one MCP per source, the agent combines results inside its context window, where mistakes happen.
You\\'re limited to whatever each provider chose to expose through its MCP server.
MCP access is all or nothing per source, with no row-level security.
Pick from hundreds of connectors: Stripe, PostHog, HubSpot, GitHub, Google and Facebook Ads, Sheets, CSVs, BigQuery and ClickHouse.
Add the tables you need to one data model, then annotate and describe them.
The agent writes SQL across sources for deep, multi-year analysis you can verify.
One MCP per source, with the agent matching records in its context window. That\\'s where hallucinations start.
Sources land in one governed data model, and the agent writes real SQL across them.
Get an agent live on the tools you already use, before you invest in a warehouse.
Proper queries across sources, not in-memory guesswork.
Expose exactly the tables and fields you want, not just what an MCP server offers.
The same row-level security rules apply to every connected source.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nNo. The Data Plane lets you start with SaaS sources directly. It also works alongside BigQuery and ClickHouse.
Use MCP for actions and supporting context, like saving to Notion. Use the Data Plane for analytics you need to trust.
Yes. Data plane tables sit in your data model, so global and custom rules apply.
Can\\'t find your answer? Get in touch.
\\nConnect your Upsolve agent to Claude with one OAuth sign-in. Ask questions in Claude and get charts and analysis from the same governed context, permissions and observability as your product and Slack deployments.
\\n \\nPasting exports into a chat assistant means ungoverned numbers and no permissions.
A second, separate agent for Claude means a second set of definitions to maintain.
Admins lose sight of what people are asking and what answers they get.
In organisation settings, choose which Upsolve agent the MCP server exposes.
Find the Upsolve AI connector in Claude, sign in through OAuth and allow the Analyze Data tool.
Ask a question in Claude and get Upsolve\\'s analysis with a chart rendered right in the conversation.
People paste exports into a chat assistant, with no definitions, permissions or audit trail.
The same Upsolve agent, context and permissions answer in Claude, and every call is logged.
Analyze Data runs the same analysis Upsolve runs everywhere else.
Visualisations render directly in the Claude conversation.
The same data model, row-level security and agent configuration as every other surface.
Every Claude call appears in the agent\\'s chat history for admins.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nThe one you pick in organisation settings under MCP agent. You can switch it at any time.
Yes. Upsolve can also be deployed as an MCP server for your end users\\' own agents.
No. Setup is a settings change in Upsolve and an OAuth connection in Claude.
Can\\'t find your answer? Get in touch.
\\nSpin up a sample project and Upsolve sets up the connection, data model, security rules, agent, charts and dashboard for you. Explore the whole platform before you connect your own data.
\\n \\nYou have to connect data, model it and write prompts before you see a single answer.
An empty workspace doesn\\'t show what good looks like, so it\\'s hard to judge the product.
When your own setup misbehaves, there\\'s no working reference to compare it against.
Choose a sample project when you create a project. Upsolve sets up the connection, data model, security rules, agent and dashboards.
Ask what data you have access to, then ask analytical questions. Answers respect the sample row-level security, so the Japan user only sees Japan\\'s athletes.
Grab the iframe or a preview link, or connect it to Claude through the Upsolve MCP. Connect your own data from project settings when you\\'re ready.
You connect data, build a model and write prompts for days before you see whether the agent can answer anything.
A complete sample workspace answers questions, enforces permissions and renders dashboards the moment it\\'s created.
Connection, data model, data security rules, agent, skills, memory and suggestions, all configured.
Query as a sample user and see tenancy working from day zero.
Favourite charts, build canvases, and open a pre-built Olympics dashboard application.
Cross-check your own project\\'s setup against a workspace that\\'s known to work.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nOlympics data plus geography and economic industry data, so you can try analytical questions straight away.
Yes. Go to project settings, open Connections, pick your data source and wire it through the data model and agent.
Yes. In organisation settings, point the MCP agent at your sample project and ask it questions from Claude.
Can\\'t find your answer? Get in touch.
\\nSave the questions your agent must get right, with the SQL you trust. Upsolve uses them as test cases and references, and scores every answer\\'s confidence before it reaches a user.
\\n \\nWords like \"worst vendor\" mean something specific at your company, and the agent has to guess.
Good answers from real users disappear into chat history instead of teaching the agent.
Every answer reaches the user looking equally sure, even when the agent reasoned its way to a guess.
Open a past conversation in observability and check the question, the SQL and the chart.
Keep the SQL as it is, or edit it first. For example, rank the worst vendors by spend only, not by order count.
Every answer gets a confidence score. Raise the threshold and low-confidence answers are held back from end users.
A vague question gets a plausible answer, and nobody can tell a verified number from a guess.
Verified questions guide the agent, and each answer carries a confidence score that can stop weak answers reaching users.
A growing library of questions your agent should answer consistently correctly.
Check the agent still gets your golden questions right after every change.
Based on matching golden questions, whether the fetched schemas and columns fit the question, and whether a good chart was produced.
All answers go through by default. Sensitive teams can set a high threshold so users only see high-quality results.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nFrom real conversations. Open a chat in observability, review the SQL and save it, editing it first if needed.
Whether the answer matched golden questions and how correct they were, whether the schemas and columns fetched fit the question, and whether a good chart was produced.
No. All answers go through by default. You choose a threshold if you want to filter weak answers out.
Can\\'t find your answer? Get in touch.
\\nData model, knowledge base, skills, golden questions, prompts, memory, chat history, playground and deploy, all under one pane of glass. Version them together and ship one release to your product.
\\n \\nThe data model, prompts, skills and examples live in different places and drift apart.
Changing one piece means guessing how it affects the rest of the agent.
Debugging means jumping between chat logs, traces and config to piece together what happened.
Data model, knowledge base, skills, assets like golden questions and charts, system prompts and per-user memory, each in its own section of the workspace.
Chat as yourself or as any user in your organisation, with a live trace of every step: golden examples, schema retrieval and chart validation.
The whole workspace is versioned as one release, with production and draft versions. Pick a version and deploy it into your product as an iframe.
Context lives in a dozen places, and a change to one piece can quietly break another.
Every piece of the agent sits in one workspace, is tested in one playground and ships as one version.
Context, monitoring, testing and deployment, all in one sidebar.
The data model, knowledge base, skills and assets ship together as a single version.
See who asked what, from which surface and at what cost, then open any session.
Test as any user and watch each step of the answer as it happens.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nThe data model, knowledge base, skills, assets such as golden questions and charts, system prompts, per-user memory, chat history, a playground and deployment.
The whole workspace is versioned as one. You keep production and draft versions and choose which one to deploy.
Yes. In the playground you can chat as yourself or as any user in your organisation and see the live trace.
Can\\'t find your answer? Get in touch.
\\nClick Make it an app, upload your logo and describe the look. An agent rebuilds your canvas as a polished HTML app you can share, embed or send as a public snapshot.
\\n \\nDefault charts on a canvas don\\'t match your product or your brand.
Turning an analysis into a polished report means a designer, a developer or a slide deck.
Sharing outside the team risks handing people live access to the underlying data.
Open a canvas and click Make it an app. Upload a logo and describe the colours, layout and design you want.
The agent adds a title, stat cards and restyled charts. Keep prompting to change or remove anything.
Share with your organisation or specific people, or create a public link and QR code for a snapshot.
An analysis lives on a plain canvas, and a polished version means rebuilding it in slides or code.
One prompt turns the canvas into a branded HTML app, ready to share as a snapshot or embed.
Your logo and design direction, applied to the title, stat cards and every chart.
Viewers can hover over values in the finished app.
Public links are snapshots. Live data needs a passkey, so nobody gets live access by accident.
Update the canvas, drill into any chart\\'s conversation, then regenerate the app with the latest numbers.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nYes. Create a public link or QR code. Public links show a snapshot, and live links need a passkey.
Yes. Your original prompt stays in the side panel, so you can keep iterating on colours, layout and content.
Work on the canvas as usual, then regenerate the app to refresh it with the latest data.
Can\\'t find your answer? Get in touch.
\\nThe Builder Agent reads your workspace and your users\\' chat history, then proposes fixes to your data model, prompts and skills. Every change lands in a draft, so production stays untouched until you promote it.
\\n \\nTables change under the hood and nobody updates the data model. That\\'s data drift.
Most tables ship without descriptions, so the agent guesses what columns mean.
The signals to fix it sit in chat history that nobody has time to read.
Ask it to describe your tables, or to review the last month of usage and suggest improvements.
It scans the database and chat history, then proposes descriptions, prompt rules, knowledge base files and skills. Approve each one, or let it run in auto mode.
Everything lands in a draft. Try it in the playground, then promote the draft to a new version for your users.
Someone has to notice the agent getting worse, dig through chats and fix the context by hand.
The Builder Agent reads real usage, proposes the fixes and stages them in a draft for you to approve.
Scans every table that isn\\'t described and writes table and column descriptions.
Finds problems like data model tables that no longer exist in the database, stale views and chart date errors.
Writes knowledge base files and generates skills from what it learns about your data.
Changes go to a draft with a full history, never straight to production.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nNo. Every change lands in a draft. You review it, test it in the playground and promote it when you\\'re happy.
Table and column descriptions, system and chart generation prompts, knowledge base files and skills.
Yes. Upload files to the knowledge base so it works from your real business rules, not only from what it reads in the database.
Can\\'t find your answer? Get in touch.
\\nStart at upsolve.ai with 2,000 free AI credits and every pro feature. Dash, our setup agent, reads your data, MCPs and files and builds your first workspace for you.
\\n \\nTrying the product means booking a demo and waiting for an account.
A new workspace starts empty, so the first day goes on boilerplate setup.
Descriptions, prompts and skills take weeks to write before the agent is any good.
Click Get started on the free tier and create an account. Connect any of 300 sources, a database, a CSV or the sample data.
Optionally connect MCP servers like Notion or HubSpot and upload files with your business rules. Dash scans every table, column, MCP and file, which takes from one to 30 minutes.
Dash writes table and column definitions, context files, skills and agent instructions into a draft. Review it, publish it, then start asking questions in the playground.
You book a demo, wait for access, then spend days on setup before the agent answers well.
You sign up yourself, and Dash builds a documented workspace with skills and instructions while you wait.
2,000 one-time AI credits and all pro features, limited only by credits.
The same logic as the Builder Agent, used to build your first workspace from zero.
300 source systems and APIs through the Data Plane, direct database connections, CSV upload or sample data.
Favourite charts, build canvases, embed in your product or invite your team to collaborate.
Upsolve builds the whole stack for data agents that stay accurate in production. Every conversation feeds the next improvement.
\\nAll 14 features of The Complete Agent Stack and every pro feature, with 2,000 one-time AI credits.
Anywhere from one to 30 minutes, depending on how much data, MCPs and files Dash has to read. In the demo it took about three and a half minutes.
No. You can start with the sample data, then connect your own sources whenever you\\'re ready.
Can\\'t find your answer? Get in touch.
\\nEverything a data agent needs before people can rely on its answers, across four layers: ontology, agent harness, distribution and the improvement loop.
\\nOne definition of ARR. Every agent answer uses it.
Read more →Data permissions your AI agent can\\'t talk its way past.
Read more →Plug Notion, Linear and more into your data agent.
Read more →A data agent that\\'s 2x faster and learns every user.
Read more →Turn agent answers into dashboards your team shares.
Read more →See every step your data agent took, and fix it.
Read more →Stripe, PostHog and HubSpot. One agent. Real SQL.
Read more →Your governed data agent, right inside Claude.
Read more →A working data agent from your first minute.
Read more →Teach your agent the right answer, and know when it\\'s unsure.
Read more →Your whole agent, in one versioned workspace.
Read more →Turn any canvas into a branded app in seconds.
Read more →An agent that improves your data agent.
Read more →Sign up free. Get a working workspace in minutes.
Read more →Each feature strengthens one layer. The improvement loop feeds back into the ontology, so the agent gets better with every conversation.
\\n