AI can make mistakes, lets verify it.
4loop the human layer.
Review AI work, checked before it moves forward.
In review
This is just one use case. Approve, reject, or send it back with a reason. You can send the review criteria with the decision, and you can add more than one decision. A recorded decision is not a promise that the draft was right.
Real AI failures
These are public cases. They show why a review step can matter.
Air Canada's chatbot said a passenger could apply a bereavement fare after he had already booked.
A tribunal found the airline responsible for the bot's answer, not just the website fine print.
Read the source story: Air Canada bereavement fare chatbotA lawyer filed a court brief with case citations that AI had invented.
The filing reached court before anyone checked that the sources were real.
Read the source story: New York lawyer AI citationsA dealership chatbot told a shopper it would sell a truck for $1, calling the reply a binding offer.
The bot stated sale terms the business never approved and could not reliably enforce.
Read the source story: Chevrolet dealer chatbot offerThe problem
Agents can draft a customer reply, a campaign, a recommendation, or an operational note in seconds. The open question is whether it is ready to send.
Without a review step, the choice is to slow the work down by hand, or to let an unchecked draft move forward. 4loop holds the draft until a person records a decision.
How it works
From the draft to the next step, a person can approve, edit, or send it back. Return is a path back to the draft, not a fifth success state.
The agent submits the draft for review.
4loop holds the draft and applies the review criteria.
An editor approves, edits, or returns it. The decision is recorded.
A recorded approval can move forward. Returned work goes back to be revised.
Return → Revise. Returned work goes back.
Check the integrity of the data and the decision through their provenance.
A draft moves through submit, route, review, and finalize. A provenance check is optional. Returned work stays held until it is revised. The record is the decision. It does not mean the reply was right.
Payload match
An agent sends a payload. 4loop matches each field to a widget, so the review shows the data. Review criteria can be sent with a decision, and a review can include more than one decision.
customerThread
You sent the blue one. I ordered the green one. Can you swap them?
codeDiff
reset.ts — the reset link still works after it expires.
documentBody
Displaying a receipt, a contract, a photo, and an ad.
checks
The name matches the account.
The amount matches the receipt.
The date is inside 14 days.
The photo shows the item.
reviewCriteria
Approve the discount when it is under 15% and the item is in stock.
decisions
Accept. The work moves forward.
Deny. The work stops, and the reason stays with it.
Refuse. The work comes back. It is not ready to decide.
Samples from different reviews. The messages are not a customer story. A field with no widget stays listed on the review.
Dashboard
We can be your center console, and the control panel for decisions being lined up. What arrived over the last two weeks sits next to the teams who can look at it. The numbers in the picture are a sample view.
Capabilities
Not a notification. Not a checkbox. Meaningful AI output can be held for a structured review before it is finalized, delivered, or acted on.
The AI turns its data into a live console view as it works, so you can see what is happening without a separate setup step.
Locked templates keep a workflow consistent and ready for approval, with room to change the parts your team is allowed to change.
4loop can match an output to a reviewer by the type of work and the risk. Routing to an outside specialist is optional. It is not included in every plan.
A review can hold more than one decision. The person can approve the next step, reject it with a reason, edit it, or return it to the agent with feedback. The review criteria can be sent with each decision. The workflow continues after those decisions are recorded.
4loop keeps the reviewer, the time, the outcome, and the reason with the output. The record is the decision.
Explore the audit ledger →Meaningful outputs can be held at the point of action until a person has reviewed them and a decision is on the record. Approval means the decision was recorded. It does not mean the reply cannot still be wrong.
Built-in generative tools, chat, and knowledge retrieval can help during review. You can add your own external API, and you can add knowledge bases for your reviews. They stay inside the same path: the output is still held until a person decides.
Explore AI tools →MCP and API
Connect an agent through the Model Context Protocol server at /mcp, or through the matching REST endpoints at /api/agent. Both routes support the same review workflow. You can send the review criteria with the decision, and you can add more than one decision.
/mcp · Streamable HTTP. An MCP client can submit work, check status, and read the decision from the same thread.
/api/agent · REST. Submit the work, sync status, and take the same recorded decision back into your app.
Agent integration
Claude Desktop, Cursor, or your own MCP client can use your 4loop API. The agent submits work for a person to review, checks status, and tracks that review without leaving the thread.
MCP tool server
An agent can submit by content type or published template, read the decision and the evaluation, answer a reviewer follow-up, and fetch the payload schema before it submits.
API and webhooks
Use the REST endpoints to submit work, sync status, and take the decision back. After a person records that decision, 4loop can also call a webhook you registered: one POST to your HTTPS address, with the payload for that call. It goes out only after the review. A failed call does not undo the decision. Remote calls are on Growth and Enterprise.
The log
4loop is meant to sit in the middle. The agent calls it, the draft is held, and the decision is written down and sent back. We can intercept your data before it is processed by an AI agent. The same path can receive data connected to future AI work. That data is hashed on the way in. The log keeps provenance, a rehash, and a stored receipt.
4loop can receive data connected to future AI work. Those bytes are hashed when they arrive: SHA-256, before that work is submitted.
Each output traces to its source, the agent, the reviewer, the change, and the reason. The original payload stays with that review.
The submitted payload is hashed on the way in: SHA-256 in a stable form. A second hash seals that digest to the review, the task, and the template.
When the work is submitted, 4loop hashes the bytes again and checks the seal. If they no longer match, the reservation is refused and the first record stays.
The decision is stored as a receipt: who decided, the outcome, and when. The receipt keeps its own SHA-256 digest. That log stays. Archive does not erase it.
4loop can then call a remote webhook and send the verified payload: one POST to the HTTPS address you registered. The call waits for the recorded decision. It does not send the unchecked draft.
A stored hash shows which bytes arrived and which bytes were reviewed. It does not mean the draft was right.
Expert review works as a callable tool inside the agent workflow. Enterprise plans describe single sign-on, routing and approval policies, reviewer pools, service-level agreements, compliance controls, and help connecting MCP or REST. That list is the published description.
AI tools
Built-in and custom AI tools can draft during review. You can add your own external API, and you can add knowledge bases for your reviews. A person still accepts, edits, or rejects what moves forward.
Review-time generation · Configurable tools · Knowledge-aware chat
Review-time generation
A reviewer can call first-party generative tools from the surfaces where they already judge the work: a bound block, the decision rail, or an iterate-with-AI step. The result can be candidate text, code, an image, video, or a document draft. It stays a preview until someone accepts it.
Configurable tools
Beyond the built-ins, you can add your own external API. A team can also attach a workspace tool or a registered executor. You choose which tools appear, which bindings are allowed, and which kinds of output a reviewer may request.
Knowledge-aware chat
A chat assistant can run on a review template so a reviewer can ask a question, compare options, or check context. You can add knowledge bases for your reviews. When retrieval is on, it can add snippets from those sources. That context stays in the conversation. It does not replace a block, and it does not approve the work.
A generated candidate is a proposal. Accepting it still goes through the review, so the person on the decision stays accountable for what moves forward.
Your stack
Add your own external API, and add knowledge bases for your reviews. OpenAI and Anthropic can be the API a review calls. Pinecone and Qdrant can be the knowledge base a review reads.
Bring the model endpoint your team already uses.
Bring the sources your review should be able to read.
These are examples a review can use. A person still decides what moves forward.
Chat
A customer conversation sits beside an advisory analysis. The panel does not send the reply.
Optional intelligent routing
You can review the work yourself. A teammate can review it. Some plans describe routing to a domain specialist. That path is optional. It is not the opening promise, and it is not included in every plan.
Who it’s for
Marketing · Devops · Cyber security · Business plan · Data · Copy · Code · Customer Service
Small business
Use AI to create more and answer customers faster, and check customer-facing drafts before they go out.
Entrepreneurs
You are the operator and the quality bar. Review the AI work that matters before it ships, and keep a record of the decision.
Startups
Give investor-facing, customer-facing, and operational drafts a repeatable checkpoint: approve, reject, edit, or return.
Everyday work
Resumes, letters, summaries, and important messages can go through a review step before you act on them.
Activity
Each line names the action and the draft. The log says what happened. It does not prove the output was correct.
Pricing
Published plan list. You can set up through an API or an MCP server. It does not open a checkout or start a trial.
Starter
Free for 3 weeks
Then $50 /mo
For individuals and small teams piloting AI review through an MCP server.
Growth
$499 /mo
Most Popular
For teams scaling AI review operations through MCP-connected workflows.
Enterprise
Custom
Talk to the team
For scale, compliance, and expert-backed AI review infrastructure.
Platform plans describe routing, records, and workflow controls. Expert verification is described as billed on use. Enterprise controls in this list are the published description, not a certification.
Frequently asked questions
4loop is a review layer for AI work. Important output can be held for a person to approve, reject, edit, or return before it is sent or acted on.
An agent submits the draft. 4loop holds it. An editor approves, edits, or returns it, and the decision is recorded. You can send the review criteria with the decision, and you can add more than one decision. You can review your own work. A teammate can review it. An outside specialist is optional.
Through an API or an MCP server. Agents can submit work, check status, and receive the decision. Availability of a live connection is not confirmed on this page.
You can add your own external API, and you can add knowledge bases for your reviews. Those stay on the review. A person still decides what moves forward.
The editor can edit the draft or return it with feedback. Returned work goes back to be revised. It does not move forward as an approval.
A person reviewing the work is the basic path. Domain expert review is described on the Enterprise plan and is not part of every plan.
This page does not start a trial or open an account. You can set up through an API or an MCP server. The prices below are the published plan list.
