The company brain, with the source attached.
Mari connects the tools where your product knowledge already lives, GitHub, Slack, Confluence, Jira, your help center, and turns them into one brain your team and its agents can ask. Answers come with citations. Facts come with an owner and a freshness date. When a source changes, Mari flags what leaned on it. It runs in your cloud, and it does one job: product knowledge.
Run it yourself, or have us run it: in your cloud, or hosted.
Apache 2.0 · 12 connectors · your cloud or ours · Mari Kit, the open-source toolkit
Today's review
6 open- DRVerify the new proration rule in the billing runbookFact check
- MGApprove the SSO onboarding guide for publishApproval
- PKRetire two stale screenshots in auth/READMEStale
A company brain with one job: product knowledge.
The record has an ontology: documents, facts, decisions, terms, and owners, linked by typed relations, each node carrying its source, its owner, and a freshness date.
12 connectors: GitHub, Slack, Google Drive, Confluence, Notion, Jira, Airtable, Asana, Dropbox, Linear, Trello, and your Zendesk help center. Drop in files too. Mari reads them in place and re-syncs as they change. Nothing to migrate, no new tool for the team to learn.
Search and chat over everything connected. Answers stream with numbered citations back to the doc, commit, or thread they came from, so a support rep or a new hire can check before they repeat it.
Facts carry an owner, a verification state, and a freshness date. Approved answers are the ones your team signed off on. Unsigned assertions show up as exactly that.
An ontology graph ties documents, commits, PRs, and decisions together with typed relations. When a source changes, Mari flags the facts that leaned on it and routes them to their owner.
@mention Mari in Slack and it answers from the record. @mention it on a GitHub pull request and it checks the change against what's known. Point an MCP client or a knowledge chat at the same record. The agent in the console can tag, approve, sync, and edit, with each action in the audit trail.
Where Mari earns its keep.
Answer from the record, not from memory
A customer asks whether the API supports it. The rep @mentions Mari in Slack and gets the answer with the doc it came from and the PR that changed it. No tap on an engineer's shoulder.
Retire the answers that stopped being true
Every support team has a pile of saved replies nobody dates. In Mari an approved answer carries an owner and a freshness date, and gets flagged when the source behind it moves.
Find out before the customer does
Eng cut the feature last sprint. The help center still promises it. Mari flags the fact and routes it to its owner before the next call, not after.
Stop being the search engine
Every "where's the doc for" lands on whoever remembers. Point people at Mari instead. It cites its sources, so nobody has to take your word for it or interrupt your afternoon.
Catch the docs the repo left behind
The repo audit scans connected repositories for documentation drift, coverage gaps, and stale pages, and offers one-click fixes that pull the missing docs in.
Answer the questionnaire from the record
Each fact carries its source, owner, and sign-off. The security review stops being archaeology: filter the registry and hand it over.
For teams whose knowledge outran their docs.
Wikis and search tools store and find what someone once wrote. They can't tell you who owns it, whether anyone checked it, or whether the source has since moved on. Mari does that part, and only that part. If any of these read like your week, it will earn its keep.
One person is the documentation
Every hard question routes to the engineer who remembers. Their answers are correct, unwritten, and unavailable the week they take off.
The same question, answered five ways
Support asks in Slack, engineering points at a Confluence page, product corrects both in a DM. Three answers, none written down, and the customer got the first one.
Nobody knows which page is current
The wiki holds six stale pages for every good one, and new hires learn to distrust all of them. Search finds all six with equal confidence.
Your AI tools answer from whatever they can reach
The coding agent, the support bot, the chat in Slack: each one repeats what it retrieved, with no idea whether it's canonical, approved, or two releases old.
What Mari doesn't do
Mari is the knowledge layer, not the whole stack. It won't write your marketing, run your help desk, replace your wiki, or take over your coding agent. It connects to those, keeps the record of what's known, who owns it, and whether it still holds, and hands the answer, with its source, to whatever asks. If your problem is a good corpus that deserves a prettier website, a publishing platform will make you happier than we will.
AI removed the cost of writing. It didn't remove the cost of being wrong.
A source, an owner, a sign‑off on each fact. Or it gets flagged.
Mari is the record of what your team knows, org-wide. Three problems it exists to stop, each with a demo below:
tap a problem below to open it
Six months on, nobody knows where a fact came from or whether the source still says it. Once every tool in the company answers questions, "says who" stops being answerable.
Mari pins each fact to the source it came from, the doc, the commit, the thread, the reviewer, and forward to everything downstream that reused it.
Lineage diagram: two source files, api/refresh.ts and sla-policy.v4, feed facts MR-4821 and MR-4825, which feed the published docs/auth page, an RFP answer, and the status page. Highlighting traces each source downstream.
- 01connect the source
- 02capture the fact
- 03check it against the source
- 04owner signs off
- 05watch for drift
A fact with no owner is just a rumor with better formatting.
One dashboard over what your team knows.
Today's review
| Task | Status |
|---|---|
| Verify the new proration rule in the billing runbook | Open |
| Approve the SSO onboarding guide for publish | Open |
| Retire two stale screenshots in auth/README | Done |
| Review the incident escalation ladder | Open |
Source pulse
Last 7 daysLive activity
Running- Docs guardrail completed a run over billing/*.md42s ago
- Dana R. refined Pricing FAQ3m ago
- Mari verified a fact in Proration rule6m ago
- Notion synced Onboarding space25m ago
- Support
You answer the same question forty times a week from a saved reply you didn't write, and learn it was wrong when the customer replies with a screenshot.
with mariAsk Mari in Slack and get the answer with its source. Approved answers carry an owner and a date, and get flagged when the product moves on.
- Product managers
You find out on the customer call that the help center promises a feature eng cut two sprints ago. You own what the product says everywhere, and none of the pages.
with mariWhen the product changes, Mari flags the pages that leaned on it and routes the fix to its owner. You stop being the human changelog.
- Developers
Every question about how it works routes to you, and the doc you'd point at was true two releases ago.
with mariMari cites the doc, the commit, and the PR, so people check the source instead of you. The repo audit finds the pages that fell behind.
- Leaders
You sign off on what the company publishes with no way to see it. Then a security review asks who approved a number, and the answer is a week of archaeology.
with mariOne dashboard over every fact's source, owner, approval, and freshness. The audit trail exists before anyone asks for it.
companies forecast to have shipped generative AI by 2026
of outbound messages forecast to be synthetically generated
AI trust, risk, and security management (AI TRiSM) is a named priority
In your cloud, on your terms.
Mari is open-source software, Apache 2.0. Run it yourself, have us run it inside your VPC, or have us host it. Your data lives where you decide.
The full security page →A Docker Compose stack with Postgres and pgvector, in your VPC. Local models through ollama are optional, and every feature has a deterministic fallback without them.
Connectors read with tokens you create and can revoke, scoped to what you choose to share. Sources stay where they are: Mari syncs, it doesn't migrate.
Each fact carries its owner, sign-off state, and freshness, and answers cite their sources. A review gets receipts, not recollections.
A custom MSA, DPA, and a security review with your team come with the paid plan.
Two open-source projects. One company brain.
Mari, the product on this page, is Apache 2.0 and free to run yourself. Mari Kit is the open-source Python toolkit for teams building a knowledge system of their own. It turns changing company sources into versioned, permission-aware, evidence-linked knowledge, and leaves the databases, the models, and the authorization policy to your application. Pick the product if you want the brain running this quarter. Pick the kit if you are building one.
Company search
Store every revision of a source, keep only the candidates a person is allowed to see, and hand back the exact evidence that is current.
Build it with the kit 02Governed knowledge
Resolve typed evidence, a text span, a record field, a table cell, into a derived fact someone can review, with a revision on every artifact.
Build it with the kit 03Agent knowledge
Turn what an agent just did into a knowledge proposal a person signs off on, instead of a memory it wrote to itself.
Build it with the kitThe questions you'd ask before you trust it.
Three ways in.
Poke the demo
The dashboard on demo data, no install. Trace a fact, fire a drift event, all from your phone.
Open the demo 02Run it yourself
Apache 2.0. Clone, compose up, connect a source. The docs walk through it, and the first cited answer takes about ten minutes. Building your own system instead? Start from Mari Kit.
Read the docs 03Talk it through
20 minutes with us on your rollout, and a full demo on your own sources if you want one.
Book a demoRun the open-source build free. Pay for the enterprise features and the operations.
Mari is Apache 2.0, and the self-hosted build is free to run yourself. The enterprise plan adds features that aren't in the open-source build and has us run it, in your cloud or ours. Pricing scales with your team, not your stack.
Self-hosted
The open-source build, on your infrastructure. Clone it, compose up, connect a source, and the record is yours to run.
- 12 connectors, cited answers, facts with owners, lineage, and drift audits
- Docker Compose, Postgres with pgvector, optional ollama for local models
- Slack bot and GitHub webhooks, self-serve
- The console and the agent, with the audit trail
- Community support (GitHub)
Bring your own cloud
Runs in your VPC. We deploy, run, upgrade, and maintain Mari inside your cloud account. You get the record without the operations.
- Everything in Self-hosted, plus the enterprise features
- Data and access stay in your account, and you can see and revoke what we reach
Hosted
Runs in an environment we operate. Your team signs in. Nothing to provision on your side.
- Everything in Self-hosted, plus the enterprise features
- Where it runs, who can reach it, and retention set out in the MSA and DPA
Both enterprise shapes come with a dedicated solutions engineer from rollout onward, onboarding for support, product, and engineering, and a custom MSA, DPA, and security review.
Custom procurement, on-prem, or a different shape entirely? Book a demo and we'll figure it out.