Your IT system,
an AI agent.
Biz AI Agent is a fully open-source AI agent framework: one business system + one DSH = one AI employee. Describe the business you need, and AI coding tools such as Claude Code and Codex build and verify the agent by following the skill that ships with it.
# 1. get the open-source repository git clone https://github.com/ITC-cloud-soft/biz-ai-agent.git cd biz-ai-agent && npm install # 2. open an AI coding tool (the skill is included) claude # or codex # 3. ask in your own words > Use the agent-builder skill to build a "travel expenses" AI agent
An agent is first an ordinary IT system.
No drag-and-drop flow editor. The business system is the system of record and flows are code, guarded by the manifest, the conformance suite, business tests, evals and logs.
The business system is the record
Data, roles, who sees what and who decides: all of it stays in the business system.
The AI only reads and proposes
It reads with the asking employee's own delegation; its only write is a proposal that runs when a person accepts it. People decide amounts.
The business system decides who is told
Recipients of notifications, approval cards and reports are decided by the business system, with the same rule as its web screens.
No personal data in events
IDs, numbers, dates and states only. Names, reasons and amounts are shown on the web after a permission check.
Permissions first
Every search and every graph answer is filtered by the asker's rights before ranking. What you cannot see, the AI cannot see.
One business = one agent
One business system and one DSH make one AI employee. The whole Biz Assistant Framework plus the agent layer (knowledge, graph, settings) in one repository.
Just describe the business.
An AI coding tool builds it.
You do not write the agent from scratch. Following the agent-builder skill in the repository, the AI builds everything from the business system to the knowledge, evals, settings screens and DSH settings, and verifies it with commands. The equipment-loans example was built from scratch by an AI with only this skill.
Claude Code
Run claude at the repository root. The skill is in .claude/skills/agent-builder/ and loads automatically; /agent-builder calls it directly.
Codex
Run codex at the repository root. The skill is in .agents/skills/agent-builder/ and is discovered automatically; $agent-builder calls it directly.
Any other AI coding tool
Ask it to "read skills/agent-builder/SKILL.md completely first and follow its steps".
Use the agent-builder skill to build a "travel expenses" AI agent: - employees ask about the expense rules and check their own claims on LINE - managers approve their team's claims; above 100,000 yen finance approves too - finance gets a monthly count report - the rules are in <your documents> (Japanese and Chinese)
What the AI does
- Asks you the decisions first, writes them into AGENT_PLAN.md and waits (with nobody to ask: the recommended default, marked "to confirm")
- Copies the reference implementation and writes the business system and its manifest, up to conformance 52 / 52
- Adds knowledge and the graph, writes the evals before tuning, and builds the settings screens, SNS approvals and reports and the DSH profile
- Verifies every step with commands and reports the results, what is not verified and what is to confirm
What you do: answer the decisions · provide the documents · review the plan and the report · try it locally · deploy
What you get is an ordinary Node.js project.
Everything is created in examples/<system-id>/ as files under git.
| AGENT_PLAN.md | the decisions and their evidence: who decided what, what is to confirm, the numbers |
| bridge.manifest.yaml | the system's self-description (ontology): roles, entities, events, the AI's tools, screen routes |
| src/app.mjs | the business system: data, sign-in, business rules (hooks), business API, settings plugin, knowledge, graph, DSH gateway |
| src/server.mjs | the entry point: port, secrets and data directory from environment variables |
| knowledge/*.md | the company documents (visibility written inside) |
| evals/ | the eval questions and their runner |
| test/*.test.mjs | tests for business rules, conformance, knowledge permissions, graph, eval floor and the settings plugin |
| dsh-profile/ | the product profile of the AI plane (DSH) |
Deploy it like any website.
Browser · IM: staff open the screens / webhooks from LINE and others
HTTPS reverse proxy: nginx, Caddy …, your domain
Business system (a Node.js web app): node src/server.mjs, with the DSH gateway
AI plane (DSH): one more process on the same machine, listening on 127.0.0.1 only
- The business system is an ordinary Node.js web app:
npm install, thenNODE_ENV=production node src/server.mjs(kept running by systemd, pm2, Docker …) - Secrets come only from environment variables; the data directory goes on persistent storage
- DSH is used only through the gateway inside the web app; IM webhooks point at the web app's public paths
- Besides that you need only a model API. A database only for pgvector
What you cannot see, the AI cannot see.
A question ── an employee asks about the leave rules on LINE
Message: dsh-im receives it and checks the employee has linked their account
Tool: the knowledge search runs with that employee's delegation
Search by rights: the business system returns only readable sections, with sources
Guards: the citation guard adds the sources, the link guard removes foreign links
Answer: back to LINE with links to the business screens
An approval ── an administrator uploads a company document
Checked for personal data; approval.requested recorded in a transaction
The approval card goes to another administrator on LINE (never the uploader)
"Approve" is decided by the same mayDecide as the web, then published
After the index is rebuilt, a count-only report reaches the report roles
Three search methods. All permission-first.
Vectors and the graph compare only the chunks the asker may read. If a model fails, search falls back to BM25F and the screen shows the error code.
| method | needs | how |
|---|---|---|
| BM25F | nothing (default) | weights title, keywords, headings and text; synonyms can be tuned |
| RAG | an embedding model | BM25F + vectors (fused by RRF), stored in pgvector or a file; the text itself is not stored |
| Knowledge graph | a base model | documents extracted into entities and relations, stored as JSON; search = BM25F + graph |
Administrators: /agent-settings
Models (keys as environment variable names only), document upload and approval, the search method, the organisation's IM bots (13 SNS + e-mail), approval and report settings. Japanese, English and Chinese; light and dark.
Employees: /me/sns
Issue a link code (6 digits, 10 minutes) and send "連携 123456" to the organisation's bot in a private chat. You see only your own links, and choose the notification types and quiet hours yourself.
The rules are set in code and configuration.
The AI reads and proposes
Write APIs answer delegations with 403. It never decides amounts.
The business system picks recipients
Each person is checked: "may they see it on the web?" If unsure, nothing is sent.
Only the person links
Through the business system's own sign-in. Administrators cannot do it for them.
No secrets anywhere
The repository, configuration, logs and chat hold environment variable names only.
Outbound settings stay closed
Opening one is an explicit change with a written reason; the checker verifies it.
Test on fictional data
Tests use fictional data and an isolated DSH home only.
Evals before tuning
20+ questions per language. hit@3 of 80 % or more, 0 leaks.
"Not verified" is said
Everything is checked by commands; what was not run is reported.
Run it on your machine first.
Node.js 24.9+ and npm, plus pnpm for the end-to-end run. No database, API key or chat account needed.
git clone https://github.com/ITC-cloud-soft/biz-ai-agent.git && cd biz-ai-agent npm install npm run vendor:install # once npm test # about 900 tests, 0 fail npm run e2e # real DSH + mock model + mock LINE: 15 / 15 PASS
Build the AI agent for your business
with AI, starting today.
All the source is public on GitHub (Apache-2.0). We also help with adoption.