Your company, queryable.
By your team and your agents.
Groundtruth is the wiki whose primary API is MCP. Policies, decisions, metrics and every agent's work live in one source of truth that people read and agents query — and every answer cites its source.
Or skip the line: start the free beta.
“A customer wants a refund after month 7. What's our policy, and who approves it?”
After month 6 customers can cancel with 60 days' notice; remaining months are prorated and refunded1. Refunds of $5,000 or more need VP Customer approval; agents quote the policy and escalate to a CSM1. This changed on Aug 12, when the 90-day notice was dropped2.
- 1Refund and cancellation policy · page · owner Sofia A. · verified 12d ago
- 2Shorten cancellation notice to 60 days · decision · accepted
Your wiki was built for humans to write and nobody to read.
Now agents do most of the reading — support, sales ops, finance close, on-call. They quote the refund policy that changed in August. They can't find the decision that explains why. They learn things nobody ever writes down.
Every team ends up building a brittle RAG pipeline over a wiki that's stale by Friday. Groundtruth replaces both.
Ask, with receipts
Every answer — in the app, over the API or via MCP — cites the exact page, decision or record it came from. No source, no answer.
Agents write back
Agents log decisions, append incident timelines and update records. Every write is attributed and badged, so humans always know who said what.
Freshness is a feature
Every page has an owner and a verified date. Stale knowledge is surfaced to humans and flagged to agents before anyone quotes it.
One URL. Your agents know the company.
Paste the MCP endpoint into Claude, Cursor or your own agent. 23 tools: ask, search, get_page, log_decision, query_records and more. Same permissions as a teammate.
claude mcp add --transport http groundtruth \
https://groundtruthhq.tech/api/mcp --header "Authorization: Bearer gt_…"{
"mcpServers": {
"groundtruth": {
"type": "http",
"url": "https://groundtruthhq.tech/api/mcp",
"headers": { "Authorization": "Bearer gt_…" }
}
}
}Four primitives. Everything queryable.
Pages
How the company works. Markdown, owned, verified. Every page has a .md twin for agents.
Decisions
Why things are the way they are. Context, options, owner, status — the ‘why’ that wikis lose.
Records
Structured facts: customers, metrics, people, vendors. Filter and sort over MCP like a database.
Activity
Every change by every human and agent, attributed. The audit trail of how your company learns.
Agents that already know your company.
Install an agent in one click. It runs every morning on your ground truth, writes its findings back, and pings you (or Slack) only when a human is needed.
Contract Agent
Finance & legalKnows every contract's cancel-by date. Paste a contract, it extracts the terms, then alerts you 60, 30, 14, 7 and 1 day before you have to give notice, and shouts if a window was missed.
Churn Agent
CustomersWatches every meeting, email, ticket and NPS score per customer. Scores churn risk with plain-English reasons and alerts you before renewal, not after the cancellation email.
Freshness Agent
KnowledgeFlags pages nobody has verified in 90 days to their owners, so your other agents never quote last year's policy.
Coming next: Invoice Chaser, Renewal Desk, Vendor Spend Watcher, Decision Scribe, Compliance Calendar.
Confluence for 2027 isn't Confluence with a chatbot.
| Legacy wiki + AI add-on | DIY RAG pipeline | Groundtruth | |
|---|---|---|---|
| Agents can read | Export / plugin | Yes, after a sync | Native MCP + REST |
| Agents can write back | No | No | Yes, attributed |
| Answers cite sources | Sometimes | If you build it | Always — or no answer |
| Knows what's stale | No | No | Owners + verified dates |
| Decisions & structured facts | Pages only | Chunks only | Decisions + records |
| Time to first agent query | Weeks | Weeks | 2 minutes |
Questions, answered (with citations).
Which agents work with Groundtruth?+
Anything that speaks MCP over HTTP — Claude (Desktop, Code, claude.ai connectors), Cursor, and agent frameworks like the Claude Agent SDK — plus anything that can call a REST endpoint.
Do you charge for agent seats?+
No. You pay for the humans on your team. Agents are unlimited on every paid plan. We think charging per agent punishes exactly the behaviour we want: putting more of your work on autopilot.
Can I import from Confluence or Notion?+
Today: paste markdown, or point an agent at your export and let it create pages over MCP. Native importers and live Slack, Drive and GitHub connectors are next on the roadmap.
Which model writes the answers, and who pays for the tokens?+
Your agents use their own model: Groundtruth hands them the cited sources and never calls a model on their behalf. In the web app, answers are written by your own Claude or OpenAI key (billed by your provider, zero markup), or stitched from the cited sentences if you don't add one. If no source matches, Groundtruth says so instead of guessing.
Is my data used to train models?+
No. Your workspace is yours. It's encrypted at rest and never used for training.
Give your agents the ground truth.
Early access is rolling out to AI-forward teams now. Join the waitlist — or start the free beta today.
Start the free beta →