Meta's consumer agent Muse works on its own on schedules and events, remembers earlier conversations, and asks on an approval card before sensitive steps. Biz AI does the same inside a company's business system, with each employee's own permissions — open source (Apache-2.0), not affiliated with Meta. The AI proposes; confirming, approving and amounts always stay with people.
Connecting an existing system still means adding things on the business side — employees and roles, a live-session check, recipient rules, approvals, reports, a personal-data check: 13 items in total. Working all of that out by hand takes time.
So Biz AI ships an integration skill for AI coding agents. This post shows how to put it into Claude Code, Codex, DeepSeek Harness and MetaCoder and let them do the integration.
The design is in an earlier post.
What is in the skill
biz-assistant-integration/
SKILL.md entry point: choosing a path, how to proceed, rules, what people decide
references/ ten procedures, read only when needed
01-assessment.md assessing an existing system (13 items)
02-new-system.md building a new system
03-existing-system.md connecting an existing one (Node.js, aliases, other languages, external mode)
04-manifest.md writing and validating the manifest
05-business-side.md hooks, events, recipients, approvals, proposals, reports
06-dsh-profile.md AI-plane profile, outbound settings, automatic tasks
07-web-and-bff.md business web screens and the BFF
08-optional-features.md manual, memory, profile, attachments, external connectors
09-rules.md MUST / SHOULD summary
10-verification.md conformance and the definition of done
templates/
integration-plan.md integration plan template
The agent reads SKILL.md first and opens only the reference it needs right now. The repository's documents are the source of truth; the skill is the route through them.
What the skill makes the agent do
It goes assess → plan → implement → verify:
| Step | What | How it is checked |
|---|---|---|
| 1 Assess and plan | read login, employees and roles, APIs, transactions, batch jobs; fill the 13 items as exists / add / alias; choose a path; write integration-plan.md
|
a person reviews the plan |
| 2 Manifest | roles and delegation ceilings, events and payload allow-lists, messages, tools, playbooks, reports |
bridge-validate manifest --strict exits 0 |
| 3 Business side | hooks (employees, authenticate, isSessionLive, recipientsFor, approvals, proposalKinds, reports, piiScan); AI writes are proposals only |
the system's own tests |
| 4 Events | the business row and emitEvent in the same transaction |
the system's own tests |
| 5 Conformance | 52 checks against a TEST environment | 52 passed, 0 known gaps, 0 skipped |
| 6 AI plane | profile, bot, outbound settings kept closed | profile checks |
| 7 Web and BFF | screens for chat linking, notification settings, approvals, AI proposals, the AI employee's profile | type checks, screens |
| 8 Automatic tasks | the delegation class and the playbooks | profile checks |
There are five paths: new on server-kit / existing Node.js with server-kit mounted / aliases on existing APIs / another language implementing Bridge v1 and proven by the conformance suite / external mode without code changes (trial).
It does not decide for you
The skill has a list of what people decide. The agent stops there, lays out the options with a recommendation, and waits — for example:
- the first business flow and what the AI may read
- which events go to whom (and which may be sent at night)
- what the AI may propose and which roles may accept it
- whose delegation automatic tasks run under
- whether any closed outbound setting is opened, and why
- which chat and which model
Rules it must keep
- never modify the framework itself (
packages/*,plugins/*,vendor/*, the profile template); fill gaps with hooks and settings - the AI's tools are GETs and proposals only; business write APIs return 403 to a delegation
- the business system decides recipients; if it can't, nothing is sent
- no personal data in events or messages (IDs, dates, counts, states only)
- no secrets in the repository, config files, logs or chat — a person enters the values
- fictional data, loopback and a TEST environment only
- never claim "it connects without changes"; the plan lists what is added
Installing it in each tool
git clone https://github.com/ITC-cloud-soft/biz-ai.git
Then copy the folder biz-ai/docs/skill/biz-assistant-integration to where your tool looks for skills:
| Tool | For you (all projects) | For the team (the business system's repo) | How it is used |
|---|---|---|---|
| Claude Code | ~/.claude/skills/ |
<repo>/.claude/skills/ |
picked automatically when the request matches the skill's description |
| Codex | ~/.agents/skills/ |
<repo>/.agents/skills/ |
automatically, or $biz-assistant-integration / /skills
|
| DeepSeek Harness |
~/.dsh/skills/ or ~/.agents/skills/
|
<repo>/.dsh/skills/ or .agents/skills/
|
picked when the request matches |
| MetaCoder | — | put it in the repo | ask it to read SKILL.md first |
# Claude Code (team)
cp -r biz-ai/docs/skill/biz-assistant-integration <repo>/.claude/skills/
# Codex and DeepSeek Harness both read .agents/skills — one copy serves both
mkdir -p <repo>/.agents/skills
cp -r biz-ai/docs/skill/biz-assistant-integration <repo>/.agents/skills/
MetaCoder (metacoder.myaiportal.net) is our AI development tool: it reads the whole codebase as a semantic knowledge graph and drives development through a 10-phase harness pipeline with machine-checked quality gates. Put the skill in the business system's repo and have it read SKILL.md first; having the whole system as a graph should suit step 1, the 13-item assessment.
How to ask
Open the agent in the business system's repository and ask in your own words:
Integrate Biz Assistant Framework into this system.
この業務システムに Bizアシスタントを導入して。最初は請求の承認から。
把 biz assistant framework 导入这个既有系统,先从巡检超期提醒开始。
Tip: start with one business flow (invoice approval, overdue inspections, …), run it next to the existing screens, then widen.
"Done" is decided by commands
The last step checks the definition of done with commands: strict manifest validation, conformance 52 / 52 (0 skipped, 0 known gaps), the system's tests, profile checks, outbound settings, type checks. Anything not checked (a real chat, a real model) must be written down as "not verified".
The conformance suite calls the system from the outside over HTTP, so you can see whether the code the agent wrote breaks a Bridge rule without reading all of it. Add an employee's name to an event payload, for instance, and the write is refused and the tests go red.
The best and most complete guide is on the official site
This post is the entrance. The integration flow, how it works, the safety commitments and how to use the skill are explained in full on the official site.
-
Official site: https://biz-ai.myaiportal.net/en/
- Integration and the skill: https://biz-ai.myaiportal.net/en/integration.html
- GitHub: https://github.com/ITC-cloud-soft/biz-ai
- Questions: GitHub Discussions
Top comments (0)