Build your AI native company.

One person can run the workload of ten. Sift gives you AI agents that work like real staff: you hire them, train them, and check their work.

Book a demo Join the waitlist
Sift Threads list with a pricing thread waiting for review
Expanded Sift retrieval trail showing three sources and a pricing contradiction
A cited pricing decision and handbook patch waiting for human approval
Q3 annual pricing2 items to approve
S
SiftAgent
⌄Searched 3 sources· 8s
Read approved pricing S2
Checked live billing S9

Use $59 annually. The approved decision and live plan agree. S2 S9

Pricing decisionReady to record
Review
Handbook patchStill staged
0 committed
A real Sift thread moves from retrieval, through a contradiction check, to a cited decision and source patch waiting for approval.

The work compounds.
The workforce does not.

12:04AM

Integrations. Choose exactly what Sift can remember.

Every connection begins with a visible boundary: the account, the selected resources, who can see them, and the import state.

Slack2 of 14 channels
#eng-billinglive
#launchlive
Direct messagesnot connected
Granola1 folder
Roadmapsynced
Private notesexcluded
Google Drive1 folder
Q3 planningsynced
Everything elsenever read
Sift memoryworkspace · scoped
Pricing decision thread#eng-billing12:04 AM
Roadmap · 14 notesRoadmap folder12:04 AM
Q3 planning · 9 docsQ3 planning12:04 AM
Scoped import complete3 sources · 0 private
Three connected sources show exactly which channels and folders are selected. Dashed routes carry only those selections into the Sift memory aperture, excluded rows never join a route, and the import ends with visible provenance.

3:04AM

Memory. One brain for everything your company knows.

Sift does not hide a conflict. It shows the records, explains how it resolved them, cites the answer, and leaves consequential changes staged.

#eng-billingApr 14 · approved

Annual is $59, flat for all new plans — approved.

Billinglive now

Active annual plans charge $59.

Handbook · §3 PricingJan 08

Annual: $49 / year, billed up front.

stale
Q3 annual pricing — fix the splitCZ · DW · Sift
CZ
You

Our docs disagree on annual pricing — $49 in the handbook, $59 on the site. Find the current price and prepare the fixes.

S
Siftagent run
⌄Searched company memory· 3 sources
Read the approved pricing decision S2
Checked live billing S9
Flagged the stale handbook figure S7

$59 flat is current. The $49 figure comes from the older Q1 handbook S7; the April decision supersedes it, and Stripe agrees. S2 S9

New decisionAnnual price is $59 flatgrounded in S2 · S9
ApproveEdit
Source patchhandbook.md · $49 → $59nothing changes until review
Review diffDiscard

Both writes staged · 0 committed until you review

Three real records — an approved Slack decision, live billing, and a stale handbook page — light up as the agent's retrieval trail reads them. The stale record is struck through and stamped, the answer arrives with citations, and both writes stay staged at zero committed.

5:41AM

Learning. Agents that get better at their job.

A correction becomes evidence, then a proposed skill revision. Replay checks the change; only an owner can make the next version active.

Investor update · draftexercise 184
S
Siftdraft

A strong quarter: usage grew 22% after the QBR and $2.4M closed against a $2.1M target.

CZ
Youcorrection

That $2.4M is projected, not closed. Signed is $1.6M — never mix the two in an investor update.

Skill · investor-updatesowner review
v2 activev3 proposed
− Report pipeline as closed revenue
+ Separate closed, signed, and projected. Cite every forward-looking number.
Replay · 3 past drafts re-checkedno regressions
Approve v3Reject
v3 active · owner approvedreversible · v2 kept
A draft claims 2.4 million closed; the owner's correction strikes it through. The correction routes into the investor-updates skill as a proposed diff, replay re-checks three past drafts, and version three only becomes active — and reversible — after owner approval.

6:58AM

Orchestration. An organization, not a task queue.

Named work moves through role-titled agents. Ownership, reporting lines, blockers, source evidence, and the return to human review remain attached.

Run · Monday launch briefgoal: a checked brief in your tray by 7:00 AM
RResearchreports to youreading 6 sources…14 facts pulled
WWriterreports to youassembling brief…6 pages drafted
✓Fact-checkreports to youverifying claims…7 cited · 1 flag
YouOwner reviewthat's youwaiting…in your tray · 6:58
Returned to Owner · 7 cited claims · 1 flagexternal actions 0
One named run crosses four role-titled lanes — research, writing, fact-checking, and owner review. The brief ticket collects its evidence at each lane, every agent reports to you, and the run ends back in your tray with zero external actions.
7:26 AM

7:58AM

Stale docs are worse than no docs.

People trust what is written. So every night Sift checks every claim your documents make against the sources of truth, and stages a fix wherever they disagree.

Truth sweep · overnight212 claims checked · 7:58 AM

Support replies within 24 hours.handbook §4

You promised Hendricks a 4-hour response.jun 30

patch staged

Onboarding fee: $150.services page

Waived for annual plans since May 12.decision

patch staged

After-hours escalations go to Dana.runbook

Dana left in March. Calls route to Sam now.people · mar 14

patch staged
209 claims agree with reality3 patches wait for your review
An overnight truth sweep checked 212 documented claims. Three conflicts surface one by one: the stale line is struck through, the cited source of truth appears beneath it, and a patch-staged stamp lands. Nothing changes without your review.

8:24AM

Your 9:00 call, already briefed.

The account's history lives in five tools. Sift assembled it while you slept: what was promised, what went wrong, what changed. Every line cited.

Hendricks & Co · renewal call9:00 AM
Renewal: $7,200 / yr, expires Jul 31two-year customer · pays annuallycrm
May complaint: two missed pickupsresolved in 6 days · they were not happymay 9
Tuesday: asked to onboard two new locationsexpansion signal · no quote sent yettue
You promised SSO by Q3 on the kickoff calljun 12 · still openjun 12
Open promise: raise the SSO date before they doflagged
In your tray, read in two minutesassembled overnight · nothing sentbriefed
A pre-call brief assembles overnight for a renewal call: renewal terms from the CRM, a May complaint from email, Tuesday's expansion request from WhatsApp, and a promise made on the kickoff call, each line cited to its source. The one open promise is flagged so you can raise it first.

8:52AM

Meetings used to evaporate.

Yesterday's calls became recorded decisions with owners and sources. The one that contradicts your pricing page was flagged, not filed.

Tuesday · pipeline review41 min · transcript
"…so let's just do fifty-nine flat for the pilot, keep it simple…"2:14
"…Hendricks kickoff, we said August fourth, right? lock that in…"2:31
"…and add two references to the proposal before it goes out…"2:48
Decision recordsaskable forever
Pilot pricing: $59 flatowner Maya · expires with the pilot2:14
Hendricks onboarding starts Aug 4owner You2:31
Proposals carry two referencesowner Sam · added to the skill2:48
$59 contradicts the pricing page ($49)held for you
Ask "what did we decide?" any morningit answers with these, cited
Three moments from Tuesday's 41-minute call transcript become decision records with owners and timestamped citations. The pricing decision that contradicts the public pricing page is held for review instead of filed. The records stay askable afterward.

9:00AM

You were asleep.
You were never outnumbered.

The rest of the day is yours.

12 minreview the night
A twelve-hour day strip from 9 AM to 9 PM. Twelve minutes of review sit at nine o'clock; the rest of the strip is highlighted open time.

FAQ

What works today, and what is the vision?

The brain and learning loop are real; the company-operating layer is the direction.

Today Sift can ingest selected sources, answer with citations in shared threads, manage reviewed and reversible skills, and connect shell-capable agents through the published CLI.

The organization, overnight handoff, and morning board on this page show how those pieces are intended to work together, not a claim that fully autonomous company operations have shipped.

What does setup involve, and which tools can I connect?

Start with one useful workflow, not the whole company.

Create a workspace, connect one or two sources, choose the exact channels, folders, labels, calendars, or repositories, and let the first bounded import finish.

The catalog includes Slack, Google Drive, Calendar, Gmail and Meet, Notion, Airtable, Attio, Folk, Granola, GitHub, Roam, and WhatsApp Business. Some update live; others poll or import on demand, so we confirm the required source and freshness before a pilot.

How do I know an answer is not hallucinated or stale?

Every answer should show its work.

Sift returns citations and timestamps, surfaces contradictions and freshness gaps, and should say when evidence is missing.

Citations make an answer inspectable, not infallible, so consequential decisions still require human review.

What can Sift see, and who can see it?

You choose the account, provider scopes, resources, and connection mode.

Personal connector imports default private; workspace connectors use workspace visibility. Customer content is not used for generalized model training unless your organization explicitly agrees.

Connected-agent authority is still being finalized, so an early-access pilot confirms whether its agents are user-scoped or workspace-wide. Read the current privacy terms.

Can Sift send messages or change other tools without approval?

Not through today's connector slice.

Current connectors read or import source material; agents can prepare and update work inside Sift with actor attribution and audit history. Skill changes use a reviewed, versioned, reversible flow.

External execution and the approval workflows shown above remain product direction until their runtime ships.

Do we have to replace our existing agents or models?

No. Sift is the shared brain, not a replacement agent stack.

The published CLI works for agents with shell access; MCP and API are additional access paths.

Hosted MCP has not yet been verified across every advertised client, and bring-your-own-model keys are not currently a self-serve commitment, so a pilot confirms the exact connection path.

How does Sift learn, and can it change itself?

It proposes changes; it does not silently promote them.

Corrections become lessons and a skill diff, replay shows which earlier cases would change, an owner approves or rejects the version, and every version can roll back.

Outcome watchers can attach verified system-of-record evidence as advisory grading; that evidence does not automatically edit skills or change resolver ranking.

Are we locked in? Can we export or self-host?

Your customer content remains yours, and Sift's canonical knowledge body is markdown-backed.

Today, export and deletion help are available by request; a complete self-serve workspace export is not yet proved in the app. A public community self-hosted edition is planned, not shipped.

For a current export or security request, email charles@sift.wiki.

What does it cost, and how do we start?

Sift is currently early access.

The public pricing page shows the intended plan ladder, but self-serve billing and checkout are not live yet.

Book a demo to confirm current access terms, provider setup, and the smallest useful pilot before connecting production data.

Your agents help you build your product.
Now let them help you run the company.

Hire one agent tonight. Connect one source, ask one real question, and review what the night shift got done — tomorrow at nine.

Book a demoJoin the waitlist
SUPPORT RESEARCH GROWTH FINANCE OPERATIONS You JUDGMENT
An org chart assembles itself: five agents — support, research, growth, finance, and operations — connect along dashed routes to a single node labeled You, whose role is judgment.