Skip to main content
For teams that hand real work to AI agents

Turn your team’s judgment into verified agent work.

Give your agents a goal. OrgX learns what your team accepts, checks every result against it, and brings you only the calls that need you.

Start freeWatch a goal become work

Free for one initiative, no card. Works inside ChatGPT, Claude and the coding agents you already use.

Goal120 assessments by 31 MarchInitiativeRapid Rehab launchPatient emailsTaskLaunch email
Markis drafting in ChatGPT · v1
Launch email
To8,412 existing patientsExample workspace
SubjectBrightline is revolutionizing recovery

Our new Rapid Rehab program gets patients back on their feet 2x faster, and it’s already helped Dana Ruiz return to marathon training.

Book a same-day assessment at any of our 14 clinics.

Book nowLearn moreFollow us

Drafting v1 · checks run before anyone sees itOpen it
Start where your team already worksAdd to ChatGPTFrom the ChatGPT app directory
Add to ClaudeCustom connector: https://mcp.useorgx.com/mcp
npx -y @useorgx/wizard@latest setupFor Claude Code, Codex, Cursor, OpenClaw and other MCP clients · All integrations

The problem

Your agents forget what your team already taught them.

A correction made in one chat never reaches the next agent or the next tool. So the same mistakes come back, and someone reads everything by hand.

  1. Monday · Maya

    Claude Codenew session

    Task: adding date-fns.

    npm install date-fns

    npm warn: package-lock.json created next to pnpm-lock.yaml

    YouWe use pnpm here. npm writes a second lockfile and breaks CI.

  2. Tuesday · Priya

    Claude Codenew session

    Task: adding zod.

    npm install date-fns

    npm warn: package-lock.json created next to pnpm-lock.yaml

    YouWe use pnpm here. npm writes a second lockfile and breaks CI.

  3. Wednesday · Sam

    Claude Codenew session

    Task: adding turbo.

    npm install date-fns

    npm warn: package-lock.json created next to pnpm-lock.yaml

    YouWe use pnpm here. npm writes a second lockfile and breaks CI.

Without OrgX. Corrected three times. Checked by hand three times.

See the work

Here’s the work. Here’s every check it had to pass.

Pick a piece of work and step through its versions. Tap any check to see what it found, how it was checked and what evidence is still missing.

Example workspace
To: Brightline patients · 8,412Subject and body · Final

Why this work exists: Help existing patients hear about Rapid Rehab and reach the booking form.Serves the goal: 120 assessments by 31 March. Accepted by a person, then judged on bookings.

A faster way back to the things you love

Preheader: Rapid Rehab is now open at all 14 Brightline clinics.

Hi {first_name},

Recovery shouldn’t take months. Rapid Rehab pairs a physio-led plan with twice-weekly check-ins. In our 2025 patient survey, 78% of Rapid Rehab patients were back to daily activity within eight weeks (n=412, Brightline Outcomes Report).

Book an assessment at any of our 14 clinics. Austin clinics can usually see you the same day. Or call (512) 555-0142.

Book your assessment

Brightline Physio · 1200 Lamar Blvd, Austin TX · Unsubscribe

Client launch email · standard v5 · 51 checksNeeds your review50 passed, 1 is not verified or not run. Standard changed v4 → v5: one check added, every other check identical to v1.
  • Links fetched
  • Renders captured in 3 clients
  • Survey figure not verified
  • Bookings read after send
1 needs your judgment6 fixed since v150 passed
Code checks26 · no model
Judged layers15 · scored apart
Skeptic6 · tries to refute
Your rules4 · from send-backs

Fifty checks or five hundred, the grid only gets denser. A check that could not run is listed as not run, never counted as a pass.

Goals in motion

Set a goal. Watch the work move it.

Agents take the tasks. OrgX checks each one and sends misses back before you see them. You approve or send back what is left, and every call you make is saved with the standard for next time.

Goal · by 31 March120 new-patient assessments booked
8 of 120 booked · read from the booking system
  • Rapid Rehab launchshown below
  • Referral program3 tasks
  • Clinic finder2 tasks
Week 1 of 6
Patient emailsMarketing
Launch emaildrafting v1
Reminder SMSdrafting v1
Ads and socialMarketing
Google ad setdrafting v1
Instagram postdrafting v1
Page designDesign
Rapid Rehab page designdrafting v1
Booking form designdrafting v1
Booking flowEngineering
Build Rapid Rehab pagewaits for the approved design
Build booking formwaits for the approved design
OrgX agrees with your team · last 20 blind picks11 of 20
Week 1Clears alone at 16 of 20Week 6
6
of 8 tasks needed a person
3h 40m
review time this week
4
sent back before anyone saw them
Learned this week

No hype words and one clear ask, from 4 send-backs. Most work still comes to a person while the judge learns.

Example workspace with illustrative numbers. Tap Approve or Send back on any card that needs you; otherwise the account lead decides. Your numbers depend on your work and how often you decide.

Across tools

Your next agent starts with what the last one learned.

A correction made in Claude becomes part of the standard. When the next task starts in Codex, the same standard and procedure load first, and the new result is checked against them.

  1. Claude · MondayTaught

    The account lead sends back the launch email: “Same-day is only true in Austin.”

  2. Standard v5Saved once

    A check, “availability scoped to where it’s true”, and the procedure for clinic copy. Its owner publishes it.

  3. Codex · ThursdayLoaded first

    The reminder SMS task starts with standard v5 and the procedure attached. Nobody pastes context.

  4. CheckedNew result, same bar

    “Austin clinics can usually see you the same day” passes the new check before anyone reads it.

Works with ChatGPT, Claude, Claude Code, Codex, Cursor, OpenClaw and any MCP client. Set up another client

How it works

Every correction makes every agent better.

Connect once. OrgX reads the work, turns what it learns into rules and skills, checks every result, starts the next run in any tool from what is decided, and hands you a receipt.

  1. 1Connect
  2. 2Learn
  3. 3Check
  4. 4Carry
  5. 5Prove

01Connect

Keep the AI tools you already use.

Claude Code, Codex, Cursor, ChatGPT, OpenClaw or any MCP client. One command connects every client on a machine. Nobody switches tools, and connecting does not start a run.

One command finds Claude Code, Codex and Cursor on this machine and connects each one.

npx -y @useorgx/wizard@latest setup

Desktop terminal, Node.js 22+. Review the changes it proposes, then sign in.

Setup guide
Every integration

02Learn

OrgX reads what happened, not the summary.

A summary says what the agent meant to do. OrgX reads every chat and session for what actually happened: each command, failure and correction. It turns each into something specific, then adds them up across sessions into rules and skills, each linked to the moment it came from.

  1. AgentTask: adding date-fns.
  2. Rannpm install date-fns
  3. Resultnpm warn: package-lock.json created next to pnpm-lock.yamlWhat broke
  4. MayaWe use pnpm here. npm writes a second lockfile and breaks CI.Issue
  5. Ranpnpm add date-fnsFix that held

Summary: “Added date-fns. Build passes.”Read: 3 sessions · 15 steps · 3 corrections

StandardMPS

Use pnpm in northstar/web; npm breaks the lockfile

Forming · 0 of 3 sessions
CheckMPS

Uses pnpm, not npm

Forming · 0 of 3 sessions
SkillMPS

Add a dependency in northstar/web

Forming · 0 of 3 sessions

Proposed until a person confirms it. Nothing OrgX extracts is approved on its own.

Fixes other teams adopted

03Check

Every result is checked before you see it.

Code, documents, designs, proposals and payment runs, each checked against your standard as soon as it is done. Start from the rules you confirmed. What misses is flagged with the reason before it reaches you, so you review what passed.

0 of 6 results checked

0 cleared0 flagged

Contributor setup guide Setup guide

  1. Clone northstar/web.
  2. Run pnpm install.
  3. Start the dev server with pnpm dev.
  4. Open a pull request.

  • Uses pnpm, not npm

    Rule: Use pnpm in northstar/web; npm breaks the lockfile

  • Judged on 5 layers, each above the bar

    Correctness and risk · Verifiability · Scope discipline · Operability · Clarity

  • Closer to work you approved than work you rejected

    6 approved and 3 rejected examples

Cleared before anyone reviewed it.

How the quality bar works

04Carry

Every AI tool starts from what is decided.

Before work starts in any connected tool, OrgX loads the rules, skills and decisions that apply, and sets the checks the result must pass. Then the agent works, and the run ends with a record of exactly what it did. Nobody pastes context.

Claude Codenext task · ~/northstar/web

Starting: adding turbo.

OrgX · loaded 3 things already decided

  • Use pnpm in northstar/web; npm breaks the lockfileRule
  • Add a dependency in northstar/webSkill
  • Launch copy leads with the agency buyerDecision, by Priya

Criteria set before work started · 3

  • Uses pnpm, not npmFrom the team’s rule
  • Follows “Add a dependency in northstar/web”Skill from 4 runs that passed CI
  • Judged on 5 layersYour standard
  1. pnpm add the package
  2. Run pnpm build
  3. Commit pnpm-lock.yaml only

Receipt · 3 of 3 criteria met · ready to approve

Ran in
Claude Code
Steps
3, listed above
Uses pnpm, not npm
Passed on the commands this run used
Follows “Add a dependency in northstar/web”
3 of 3 steps done
Judged on 5 layers
Each layer above the bar

Also inOpenClawOpenCodeAny MCP client

05Prove

Receipts decide what reaches you.

Every run ends with a receipt: what was asked, the criteria set before it started, and the proof for each. Complete receipts are ready to approve in one read. A receipt missing proof lands with the person who owns it, with the reason. That record is how you review less, safely.

Real evaluation records

Range

Any work your team can judge, OrgX can check.

Code is the easy case. Emails, proposals, designs, runbooks and analyses get the same treatment: checks by code where a rule is exact, a calibrated judge where taste matters, and results read after it ships.

A faster way back to the things you love

78% back to daily activity in eight weeks (2025 Outcomes Report)

source cited

Also: Launch emails, Landing pages, Case studies, Social posts, Ads, Blog posts, Newsletters

Reads from Mailchimp, Notion, Slack

  • Every number cites a sourceCode check
  • People and customers named have a releaseCode check
  • Voice matches approved workJudged
  • One clear askJudged
  • Urgency and claims can be defendedSkeptic
  • Bookings or clicks inside the range you setResult after it ships
Direction comes from
  • Linear
  • Notion
  • Google Workspace
  • Slack
Agents work in
  • ChatGPT
  • Claude
  • Claude Code
  • Codex
  • Cursor
  • OpenClaw
  • Figma
Results are read from
  • Mailchimp
  • Stripe
  • HubSpot
  • GitHub

Why you can trust the checker

The checker has to agree with you before it decides anything.

Agreement is measured on blind picks against your own decisions. It decides which results can clear without you. It never grants an agent new permissions.

Agreement decides what reaches youMeasured, per standard

16 of 20 blind picks agree with your team

The line
16 of the last 20, after at least 10 picks. Below it, every result goes to a person.
Blind
You pick between two versions before you see the judge’s pick. “Both need work” and “Can’t tell” don’t count either way.
Different family
The judge comes from a different model family than the agent that did the work.
Escapes
A mistake found after acceptance sends that kind of work back to a person until agreement recovers.

Example standard. Your line is set per standard.

Permissions decide what agents may doSet by an owner, never by a score
  • Draft, research, edit on a branchAgent
  • Open a pull request or a draft for reviewAgent, after checks
  • Send, deploy or spend through OrgXA person
  • Accept its own workNever

High agreement means fewer results need your review. It never lets an agent send, spend or deploy on its own.

After it ships

Finished work isn’t the finish line. Results are.

OrgX reads results back from the tools you connect, compares them with the goal, and proposes the next move for a person to approve.

Goal · 120 assessments by 31 March

86 of 120 booked, 71 of them from the launch email. Pace needed today: 96.

  1. DeliveredLaunch email accepted

    Accepted by the account lead on its receipt, with 51 checks passed and 1 not verified.

  2. Read backGoal behind pace

    86 booked against 96 needed by now. Read from the booking system, not reported by an agent.

  3. Observed41% leave at the insurance step

    Visitors from the email reach the booking form, then drop at the question about insurance.

  4. Proposed next moveMove insurance after the time slot

    A task for Eli, with the booking-flow standard attached. It starts when a person approves it.

Example workspace with illustrative numbers.

What you get

Check less. Explain once. Pay once.

  • Quality

    Checked before it reaches you

    Every result is checked against your standard first. What misses goes back to the agent with the reason, and what needs judgment reaches the person who owns it.

  • Speed

    Explain it once

    A correction confirmed once is in every later session, in every tool your team uses. Nobody pastes context or re-briefs a new agent.

  • Savings

    Stop paying for the same mistake

    In 3,307 of our own agent sessions, 140 separate sessions hit the same failure, each starting from zero, and more than 1,100 calls went to failures an earlier session had already hit. Each one is a run paid for twice.

    Source: Trail on OrgX’s own Claude Code and Codex history, September 2026
For you
Approve on a receipt instead of redoing the check.
For your team
A new teammate’s first session already follows the team’s rules.
For your clients
The proof travels with the work.

Start free today

  • Free for one initiative, no card.
  • One command connects the tools you already use.
  • Team plan: If your 60-day proof report shows no measured improvement, we refund the plan.
Start free

For agencies and client work

Show clients the work,and the proof behind it.

Send one link instead of a status call. Your client sees what shipped, who accepted it and why, with every receipt attached.

  • Sign-off happens on the evidence, not on a meeting.
  • An invoice or renewal comes with the record that backs it.
  • Your brand stays in front of your client.
Book a closeout review
An OrgX proof room for an example client: three delivered artifacts, the launch narrative preview and its acceptance record.

You stay in control.

Send, deploy or spend
OrgX-managed sends, deploys and spend stop for a person.
Your keys
Your workspace connects with keys you own and can revoke.
Decisions keep their owner
Every recorded decision keeps who made it and the evidence behind it.
Security and access
Pricing

Choose how you put agents to work.

Try one initiative for free. Run ongoing work with Starter. Share client delivery and its evidence with Agency.

Compare plans

  • Free

    $0Forever · no card

    Try your first initiative with all 7 agents.

    • One active initiative with its work record
    • Manual guidance before consequential work moves
    • $2 of agent usage each month

  • StarterFor your own work

    $98/month · billed monthly

    Run ongoing work with approvals and evidence attached.

    • Unlimited initiatives with durable work records
    • Sends, deploys, and spend stop for your approval
    • Decisions, artifacts, and receipts stay connected
    • $10 of agent usage each month

  • Agency

    $298/month · billed monthly

    Show clients the work, its evidence, and who accepted it.

    • Co-branded proof rooms clients can inspect
    • Accepted deliverables keep their evidence and next action
    • Up to 5 seats for client-delivery teams
    • $40 of agent usage each month

  • Enterprise

    CustomScoped with our team · seat-based or annual

    Answer to governance teams with the record intact.

    • Enterprise identity, audit trails, and data controls
    • Dedicated capacity, SLAs, and executive reporting
    • Unlimited agent capacity

Want OrgX to set up your first client loop? Client Proof Loop, $2,500/month. First proof room within 7 days.See the managed offer

What’s new

Research, releases, and field notes

All posts
  1. · Product updateOrgX is in the ChatGPT plugin directoryConnect your OrgX workspace to ChatGPT and start with the decisions and project context your team has already recorded.
  2. · GuideAI Agent Handoff Template: Keep Decisions, Evidence, and ApprovalA reusable handoff template and a fresh-session test for carrying decisions, rationale, owners, artifact versions, and permission boundaries between AI agents.
  3. · GuideClient Closeout Checklist for AI-Delivered WorkA practical closeout packet for AI agencies: scope, versioned deliverables, evidence, exceptions, named acceptance, and the next action.
Questions

Before you start

Still deciding? Bring one piece of agent work to a 30-minute review.

Book a proof review
What does OrgX do?

OrgX is proof for AI-delivered work. Claude Code, Codex, Cursor, and OpenClaw write to one shared record, so a decision keeps its provenance (the sessions, the clients, the evidence) no matter which agent picks the work up next. Capture depth still differs by client today.

Can we control what agents are allowed to do?

Yes. Each initiative runs in manual, gated, or autopilot mode, and each connector is set to allow, ask, or deny. Sending, spending, deploying, merging, and deleting always stop for a person in every mode. Checking finished work against a published quality bar is in beta.

Do we have to switch AI tools?

No. OrgX connects through MCP and plugins to Claude Code, Codex, Cursor, ChatGPT, and OpenClaw. Setup differs by client; the record is shared across all of them.

What counts as “accepted” work?

An agent reporting done is a claim, not an approval. An agent can never approve its own work: that floor is enforced in code and fails closed. Work counts as accepted when an owner records the ruling.

Can someone outside OrgX verify the work?

Agent Work Receipt v0.1 is an open, account-free format. A receipt carries its evidence and integrity hash, and it can be checked with the public validator without an OrgX login.

What does it cost to start?

The Free plan covers one active initiative. Paid plans start at $98 a month. If you want us to run it with you first, two scoped episodes on one of your real deliverables cost $250.