Private preview · permissions armed · fail-closed

Your company. Under your command.

Every source you already have, unified into one knowledge graph that knows who may see what and what was true when. Ask anything in plain English; the answer comes back cited, trimmed to your permissions before the model sees a token, and correct for the date you asked about.

Request private preview Watch it answer ↓

No card required · connects in 2 minutes · your data stays yours

  • ∞Sources
  • EveryAnswer cited
  • 0Sources leaked
  • as_ofAny date
app.slashh.io/ask
dana@
Actual product, July 2026. Sample data — plausible internal content, not a customer's.

Eight connectors over OAuth

You approve on their site. Two minutes.

  • Slack
  • Notion
  • GitHub
  • Google Workspace
  • Linear
  • AWS
  • GCP
  • Azure
  • …or describe any other tool

slashh stores the token encrypted and pulls the first batch straight away — then re-syncs on its own cadence: cloud accounts every 6 hours, everything else every 2. Queued, rate-limited, backed off when a credential breaks. Nobody clicks refresh.

Anything else? Describe it in plain English and slashh proposes the REST or MCP wiring for you to confirm.

The whole product in one interaction

Ask your company anything. Get an answer you can check.

Retrieval is trimmed to what you may see, the answer streams back with every asserted fact bound to a source, and you can open any of them. Hover a citation chip to see the sentence it came from.

Live in this page — the same shape as the real answer, running from a small fixture

asked as dana@acme.com · finance, leadership as_of today retrieval trimmed before the model sees a token

Ask something else

One diagram, five stages

Eight systems in. One cited answer out.

Keep scrolling. The diagram builds itself as the notes explain each stage — the eight sources, the edges pulling inward, the graph where every fact carries two clocks, the permission membrane closing, and finally one question in and one cited answer out.

01 Sources

Eight systems, and no shared map between them

Slack, Notion, GitHub, Google Workspace, Linear, AWS, GCP and Azure. Each one holds part of every answer your company has already worked out; none of them holds the whole thing, and nothing can query across them. Anything not on the list: describe it in plain English and slashh proposes the REST or MCP wiring for you to confirm.

  • Slack
  • Notion
  • GitHub
  • Google Workspace
  • Linear
  • AWS
  • GCP
  • Azure

8 connectors over OAuth · you approve on their site · tokens stored encrypted

02 Edges

Everything pulls toward one place — and keeps pulling

You authorize once. After that a scheduler re-syncs every connection on its own cadence: cloud accounts every 6 hours, everything else every 2. Queued, rate-limited, backed off when a credential breaks. Nobody clicks refresh, and nothing quietly goes stale in the corner.

Re-syncs every 2h · cloud every 6h · self-updating by design

03 Graph

It becomes a graph, and every fact carries two clocks

Each document is read into one graph: the people, decisions and topics inside it, and how they connect. Every fact records when it was true and when we learned it — so “what did we know in March?” is a question with a correct, cited answer. Facts are invalidated, never deleted.

Bi-temporal · as_of on every read · superseded facts kept and flagged

04 Membrane

A permission membrane closes, and it fails closed

Retrieval is trimmed to owner-ACL ∩ granted-scopes before the model gets a single token. Anything outside it is not summarised, not hinted at, not counted — it is never retrieved. An agent inherits its human's ceiling: it can be given less, never more.

Fail-closed · trimmed before generation · deniedByAcl is logged, not leaked

05 Answer

A question goes in. A cited answer comes out.

Ask in plain English and the answer arrives with its provenance attached — every asserted fact carries a source you can open. If the answer leans on something since superseded, it is flagged with the current fact. Agents get exactly the same treatment over MCP, and every call lands in an append-only, hash-chained log.

Every answer cited · freshness guardrail · runs offline with no API key

The payoff

Eight systems went in. One answer came out, and you can check every clause of it.

That is the whole product in one sentence. The eight sources are never merged into a pile of text — they become one graph where every fact knows who may read it and when it was true, so a single question can cross Slack, Notion, GitHub, Google Workspace, Linear, AWS, GCP and Azure at once and still come back with a citation per claim and nothing you were not allowed to see.

  • 8 systems queriedone graph, one query
  • 1 cited answerprovenance per claim
  • 2 clocks per factvalid time · learned time
  • 0 sources leakedtrimmed before the model
  • Every call loggedappend-only, hash-chained

Stage four, demonstrated

Same question. Two people. Two answers.

Watch the ACL gate sweep down the candidate rows: each one flips to permitted or withheld before anything is generated. Hover or focus any row to see the rule that produced the decision — including the ones that denied it.

one question · asked twice · nothing else changed

—

right-hand asker

Being honest about this mockup: in the product, the right-hand asker sees the answer and a withheld count — never the hatched rows. We are showing them here because the mechanism is the point. A real denial returns nothing at all: no title, no snippet, no count of what lives behind it beyond the single number the asker is told.

visibility is strictly nested: agent:brief-bot ⊆ Sam ⊂ Dana — an agent inherits its human's groups and holds a narrower scope grant, so both operations only ever narrow the set.

Four workflows

The jobs this was actually built for.

Four scenarios, drawn from the jobs the product is built for. Illustrated, not customers.

Priya Raman

Product engineer · day three

Three days in, half the words in her team's channel are codenames. Nobody writes a glossary for the language they already speak.

What she touches
Ask · Jargon Decoder · Entity Profiles · Decision Log

app.slashh.io/ask
priya@
The slashh Ask page answering a question, with citation chips inline in the answer and an entity profile beside it.
Ask · the question typed the way she would say it out loud.
Real capture of the slashh portal, July 2026. Sample data.

The point: onboarding context she would otherwise have to interrupt four people to assemble.

Marcus Adeyemi

Site reliability · on-call · 02:14

Billing is paging and he has never touched this service. The people who built it are asleep in another timezone.

What he touches
Search · Who-Knows-About-X · Decision Log · Contradiction Inbox

app.slashh.io/graph
marcus@
The slashh graph explorer: people, services and decisions linked as nodes, with a detail panel for the selected node.
Search across three systems at once, already trimmed to his permissions.
Real capture of the slashh portal, July 2026. Sample data.

The point: the answer existed at 02:14. Nobody could find it.

Lena Vogel

Deputy general counsel · diligence

An acquirer's counsel asks what an agent was able to see in March — and asks her to prove it.

What she touches
Org Time Machine · Audit Log · Compliance Evidence Export

app.slashh.io/contradictions
lena@
The slashh contradiction inbox: two incompatible facts shown side by side with the source, the date each was learned, and a superseded flag.
The graph as it stood on 1 March — the same read the agent got, replayed.
Real capture of the slashh portal, July 2026. Sample data.

The point: this is the question no wiki and no RAG bot can answer at all.

Tom Ishikawa

Engineering manager · someone leaves Friday

Eleven years of undocumented context walks out in four days, on a week when nobody has time for a handover.

What he touches
Offboarding Brief · Entity Profiles · Who-Knows-About-X · Audit Log

app.slashh.io/command-center
tom@
The slashh Command Center: populated KPI tiles, sync status for all eight connectors, and a panel showing where the money is moving.
The brief assembles from what she actually touched, not from what she remembered to write down.
Real capture of the slashh portal, July 2026. Sample data.

The point: a handover in an afternoon instead of three weeks of meetings.

Inside the product

Every claim on this page sits next to the screen that produces it.

Four real captures of the portal as it stands in July 2026. The data in them is plausible demo content — sample names, sample figures — because the screens are ours and the workspace is not yours yet.

01 · Command Center

Everything that moved while you were not looking.

One screen for the state of the brain: which connections synced and when, what changed, what contradicts something else, and where the money is moving. Every tile is a query into the same graph, so every number can be opened down to its sources.

  • Self-updating. The scheduler re-syncs on its own cadence — cloud accounts every 6 hours, everything else every 2. Nobody clicks refresh.
  • Broken credentials are visible, not silent. A connection that backs off says so on the front page.
  • Nothing here is a summary of something you may not read. The tiles obey the same fail-closed retrieval the answers do.
app.slashh.io/command-center
dana@
The slashh Command Center with populated KPI tiles, connector sync status, and a where-the-money-is-moving panel.
Command Center. Real capture, July 2026. Sample data.

02 · Graph explorer

One graph, not eight piles of text.

People, teams, services, decisions, documents and topics, with the edges the documents themselves implied. Walk it when you do not yet know what to ask — the shape of the neighbourhood usually tells you.

  • Every node is an entity profile — assembled from every source at once, not from one system's idea of a record.
  • Every edge carries its evidence. Click one and you get the thread, the commit or the page that produced it.
  • What you cannot see is not drawn. The graph you are shown is already the intersection of your ACLs and scopes.
app.slashh.io/graph
dana@
The slashh graph explorer showing linked people, decisions and documents, with a side panel for the selected entity.
Graph explorer. Real capture, July 2026. Sample data.

03 · Value & ROI

We label the models as models.

There is a value screen, because somebody always has to justify the line item. Every figure on it carries a provenance tag: real means measured in your workspace, modelled means derived from a published figure and an assumption we show you.

  • The honesty is the feature. A dashboard that cannot tell you which numbers it invented is not evidence.
  • The published starting point: knowledge workers spend roughly a day each week hunting for information that already exists. Your recovery is modelled from your own query volume.
  • Counts that are real are real: questions asked, sources returned, sources withheld, contradictions filed, calls logged.
app.slashh.io/value
dana@
The slashh Value and ROI screen, where each figure carries a MODELLED or REAL provenance tag next to it.
Value & ROI — note the MODELLED / REAL tags. Real capture, July 2026. Sample data.

04 · Agents over MCP

The same brain, over one config block.

There is a desktop surface for people, and a stdio MCP server for their agents. Both read the same graph through the same membrane: trimmed first, cited always, logged either way. Thirteen tools are reachable over MCP.

  • Agents act on behalf of a human. Keys are hashed; every call is appended to a hash-chained log with the principal, the on-behalf-of and every ACL decision.
  • LangGraph and CrewAI adapters, or take the SDK straight: npm i @slashh/core.
  • Runs offline with no API key — deterministic embedder, extractive answers, still cited. Self-hostable and air-gappable on an MIT-licensed open core.
  • memory_search
  • memory_ask
  • entity_profile
  • who_knows
  • decisions
  • glossary_define
  • memory_write
  • + 6 more · 13 total
claude_desktop_config.json 
{
  "mcpServers": {
    "slashh": {
      "command": "npx",
      "args": ["-y", "@slashh/mcp"],
      "env": { "SLASHH_TOKEN": "sk_live_…" }
    }
  }
}
13 tools · stdio · no gateway
slashh desktop
stdio · MCP
The dark slashh desktop surface, headlined “Your whole company, one question away”, with an ask field and recent cited answers.
The desktop surface. Real capture, July 2026. Sample data.

Honest comparison

The “No”s are part of the pitch.

Every one of these categories is good at what it was built for. The fastest way to see what that is, is to keep the honest answers in the table.

Capability slashh Enterprise search Wiki RAG bot
What it isThe company brainAn index of documentsPages people maintainA similarity index
Cites its sourceEvery answerLinks onlyManuallySometimes
Knows who may see whatFail-closed ACLsPer-connectorManualRarely
Knows what was true whenBi-temporalNoPage historyNo
Flags contradictionsBuilt-in inboxNoNoNo
Keeps itself freshSelf-evolvingRe-indexHumansRe-embed

Open core → Pro → Enterprise → Cloud (metered). The core is MIT-licensed and self-hostable, so the exit is always open. SOC 2 evidence room and control mapping with CSV export for auditors — certification is on the roadmap, and we label roadmap items as roadmap.

Private preview

Connect one source. Ask the question you gave up on.

Open to a small cohort at a time, so we can wire your odd internal system properly instead of shipping you a wrapper. Two minutes to connect, and the first answer arrives cited.

Request private preview Watch the diagram again

No card required · connects in 2 minutes · your data stays yours

MOCKUP · COMBINED 02+03+04+12