Default is Codex Local. Add an endpoint now, or save and configure it later.
Copilot
Preview first
Copilot needs an AI model
Hello, how can I help?
Drop an Excel or CSV file into this chat, or type what you want done. I'll summarise it and show you a plan first. Nothing is changed until you approve.
Copilot
Ready.
Copilot
Drop the file to attach it to the chatDrop .xlsx, .csv, .tsv or .txt — the file is only attached; the review starts when you press Send
Project Management
Projects
Auto-refresh
Not sure what to do next?Start with Copilot to review a file, create the project, attach a blueprint, and run a small QA sample.
Drop files, a folder, or a .zip here, or click to browseTables become file tasks. HTML/Markdown become a document set you can attach to a dataset view.
Which blueprint?
Options
Files wait here until you press Add file task(s); the settings below then apply. Leave this off and files upload as soon as you add them, without starting a run.
Full run is available for one file at a time; multiple files are always queued as separate safe test runs.
Blueprints available in this project
Choose the blueprints this project can use, then add or run files without going back to chat.
Add the blueprints you need, then add new files or run existing ones. Copilot (the AI assistant) is optional, not required.
Project details
Name, description and the language generated content is written in
Project team
Share this project with users in the same organization
Advanced output/review metadata
Extra fields for tracking inside the project. They are not the main Excel headers and need no change before a run.
optional
Generated CSV/XLSX files keep the original Excel headers first and foremost. The fields here are extra metadata for review and internal use only.
Delete this project
Permanently removes this project's files, document sets and dataset views. Runs are kept for the record. You'll see exactly what goes before anything is deleted.
Your file keeps its own headers. We read your source columns and write results only into the output columns you mark. If you mark none, we add the default output columns at the end. CSV and XLSX run downloads end with a Row status column.
Or create a new project
Column Mapping
Column Roles AI writes only into columns marked Output — Read and Reference columns are never written, and Ignore columns are left alone
Baseline locale is the locale preserved 1:1 as the content baseline — not the output language. It picks the uploaded WYSIWYG column when the file has one, and on a blueprint that reads an HTML archive it picks which locale’s archive is the baseline; the other locales stay evidence. Saved with the mapping, so a file carrying no locale columns keeps the choice. Output language (this run) applies to this run only; it is not saved to the mapping or the project.
Series Code
Official Series Name
Catalog Description
Draft Blueprint
Step 2
Choose processing blueprint
The blueprint defines stages, QA rules, and the quality target for this run.
AI Blueprint BuilderDraft first
Describe the blueprint, column intent, review rules, or output you need. AI creates a validated blueprint draft; it cannot run arbitrary code or change uploaded headers.
Run mode
Column mapping for this run
Sample page check
When this switch is on, a test run first opens one sample product page from your file to check it can be read. If the check doesn't pass, the test run doesn't start. Full runs skip this check.
Type a command to create a new blueprint, or press Open in the list below to edit one visually, then use it by adding it to a project.
Policy (JSON) — truth-gate rules, table rules and market rules for this blueprintDefault
Using the default policy (leave empty to use the system default)
Controls whether a node whose capability is "Input evaluation" really runs — if Off, no AI is called at all (zero cost) even when that node is in the blueprint
Controls whether the verified tables / reference data (TH/JP/KR/US WYSIWYG) are always put back into the generated content. With "AI discretion", a task written in the "Generate E/F/G" node (for example "no tables") takes real effect, but the evidence-coverage score drops accordingly if the reference data goes missing — nothing is hidden
Only used when the blueprint has an "Archive index / Series identity match / Archive evidence" node. Without those nodes, what you enter here is never read · The file is a plain .zip placed on HTTPS, no password needed, and the system fetches only the files it needs, not the whole archive
You can add several languages. The system goes through them from top to bottom — a series that the first language's archive does not have, or has only as images with no text, is looked up in the next language instead of being dropped at once as "no source" · Language codes useTH / JP / KR / USas in the market profile — typing them as the file names spell them (KOR, USA, THA) works too, and the system converts them and says so in the run log · Every locale in this list is used as an automatic reference source, added after the order set in the market profile, from top to bottom of this list · The baseline is not added twice · Turn a single row off with its "Use as reference" box
Measured on a real archive on 2026-08-23: in an 81-row test file, 14 rows (17%) have source text longer than 4,000 characters, the longest 15,525 — if you leave the default, the part beyond the limit is cut off
Rules · defaults for every node in this blueprint
Set the mode of blueprint-level rules here. Each node can override them in the Edit node panel. Changes are written into Policy (JSON) and saved with the blueprint's Save button
Edit raw JSON
Leave empty to use the default policy. If you enter only part of it, the system merges it into the existing policy for you — no need to retype every field. When the AI drafts a policy, the text appears in this box for you to check first. Nothing takes effect until you press Save
Run settings — batch size, quality target, scoring weights and export formats for this blueprintDefault
These values are used when a run is created from this blueprint. Leave a box empty to use the system default
Export formats
Tick none to use all the defaults (HTML, CSV, XLSX, ZIP). Choosing HTML always brings a ZIP along
Research
Normally the workspace's research setting always wins over the one set in the blueprint whenever the workspace has an adapter set. Tick this box to make this blueprint use its own setting (offline / local AI / API search) regardless of the workspace — the API key is still the workspace's, as before
What to check in this blueprint
This blueprint has no stages yet
Blueprint Governance
Decision records
None yet
These records explain changes to this blueprint only. They do not affect the project's Decision Log
Blueprint Governance
Revision history
None yet
Every version of this blueprint, newest first — read-only: you cannot roll back or edit from this page
Execution
Run History
Auto-refresh
System Trace
Audit Log
Use this to compare operator/AI intent against backend execution. Run logs show Task ID, Run ID, and the current execution mode explicitly.
Row Review
Run rows
This file has a newer run. You are checking an earlier one.
HTML Preview
Generated HTML
Choose a row to inspect the complete customer-facing HTML and its QA explanation.
Generated customer outputs
Rendered customer HTML
View complete HTML source
#
Series
Status
Score
Decision
Reason
Actions
Release notes
What's new
What changed in MISUMI AI Automation, newest first.
Organization
Organization
Everything a run in this organization uses: its generator model, its provider API keys, and its research, QA and evaluator policy.
Used by this organization's runs, Copilot and its other AI work while Codex Local is the generator. A research step that calls Codex Local under a different generator keeps the adapter's own default model.
Codex Local sign-in (ChatGPT)Loading…
ChatGPT sign-in for this organization is managed by an admin.
Set the Codex local endpoint in Settings › Advanced first — sign-in needs it configured.
Technical details
Cancel the pending sign-in first
Go to the link below and enter this code to finish signing in.
Checked between batches, so a long run can go a little past the limit before it pauses.
Daily limits are set by an admin.
Managed by your organization admin or manager.
Generator model
Codex model
ChatGPT sign-in for this organization is managed by an admin.
Provider API keys
Daily usage limit
Checked between batches, so a long run can go a little past the limit before it pauses.
Daily limits are set by an admin.
Organization
Research & QA
Research (Serper) key
Evaluator key
Research / Search layer
Choose whether research runs from uploaded input only, the selected AI/local endpoint, or an external search API before generation.
This mode uses matching MISUMI catalog pages as the technical basis, then uses SERP for buyer language and FAQ intent. It still blocks unsafe HTML and invented exact specs, standards, safety, or commerce claims.
Semantic AI QA
Enabled by default: after generation, the same real provider/model reviews hallucination risk, unsupported claims, and buyer usefulness. Deterministic truth and HTML blockers remain the final authority.
Evaluator layer
Scores Columns H/I after generation. Use the local heuristic now, or route scoring to your own API without tying the workflow to Claude or Serper.
The evaluator API key is set above, under Evaluator key. The endpoint it calls is a deployment setting an admin controls.
Run performance
Run speed
Loading…
Loading concurrency settings…
New values apply to runs started from now on — runs already queued or processing keep the settings they were created with.
Technical details (admin)
Environment default for rows in parallel per run:
Environment default for runs in parallel:
Codex Local's own adapter concurrency (CODEX_EXEC_CONCURRENCY, CODEX_LOCAL_RUN_CONCURRENCY) is set by the deployment's container environment, not by these fields — its adapter processes requests in a single queue, so raising the values above will not speed up Codex Local runs.
Rows in parallel per run
How many rows inside one run the server generates at the same time.
Runs in parallel
How many separate runs — across projects and users — the server executes at the same time.
Only a platform admin can change concurrency, because it sets how much of this server every organization shares. Ask an admin if it needs to change.
Technical details (admin)
Codex Local's own adapter concurrency (CODEX_EXEC_CONCURRENCY, CODEX_LOCAL_RUN_CONCURRENCY) is set by the deployment's container environment at deploy time and is not editable here.
Admin
Create a user account
Admin only
Email
Role
Status
Action
Admin
Organizations & teams
Team access
AManagerruns this organization: its projects, its blueprints, and its provider API keys.AMemberbelongs to it but manages nothing until they are added to a project. Only an app admin can appoint or remove a manager.
Account
Password
Protected
Account
Language
Choose the language of menus, buttons and labels on this screen. It is remembered on this browser. Uploaded files, exports and generated content are never translated.
Admin
System AI Configuration
Admin only
Settings that belong to this machine rather than to any one organization: the endpoints this server sends content to. Research, semantic QA and evaluator policy are chosen per organization — and an organization that has not chosen its own follows the values here.
Sign-in to ChatGPT is per organization — see Settings › AI & keys.
Evaluator endpoint
Where this server posts row content for scoring. It stays a deployment setting on purpose: a per-organization endpoint would be a way to send customer content somewhere the platform admin cannot see. The evaluator'smode and timeoutare chosen per organization, under Settings › Research & QA.
Runtime
Agent Map
Operations
Queue & QA Monitor
Admin only
Runtime Skill Registry
A skill is only a versioned set of settings, not code: pick the capabilities the system offers, define the input and output shape and the limits, then press Publish. Do not enter code, commands, links, tools or custom handlers.
This system cannot connect to the internetat all.A published skill is still only configuration data; the system does not yet run the code written in a skill.
Cross-organization
Blueprint sharing requests
Blueprints you can request
This list shows only blueprints whose owning organization has turned on "shareable" — you see just the name, the owning organization and the last-modified date, not the stages, policy, criteria or output columns until the request is approved.
Incoming requests
Requests for blueprints owned by your organization — once approved, the requester gets a separate copy to edit on their own. Rejecting requires a reason and locks the request until an admin reopens it; a rejection cannot be undone.
Outgoing requests
Requests you or your organization have sent — you can withdraw a request that is still waiting. A request with no decision after 30 days expires on its own, with nothing for you to do.