Agent-first command tree
From connections and datasets to HQL, dashboards, elements, and authorization, one hbi command tree covers the main BI engineering chain.
Execution layer
Let AI agents execute BI engineering from the terminal
HENGSHI CLI, surfaced through the hbi command, turns data access, HQL queries, dashboard generation, permissions, and delivery operations into a tool surface agents can call directly, together with repo-managed hbi-* skills. Whether the runtime is a coding agent like Claude Code or Codex, or a persistent agent like OpenClaw or Hermes Agent, the workflow moves from "telling people where to click" to "executing the work for you."
Repo-managed skills
The repo ships with hbi-* and hql-expert skills that turn data, modeling, dashboarding, permissions, and orchestration into reusable runbooks for agents.
Engineering-grade security
Tokens stay in the system keyring with OAuth and enterprise SSO support, instead of leaking into local config files.
Structured output
Native JSON, YAML, and table output reduces the prompt and context budget agents would otherwise spend on parsing.
From context discovery to asset generation, the agent no longer has to stitch together loose APIs or invent its own BI runbook.
$ hbi auth status --output json
{ "status": "authenticated", "provider": "oauth" }
$ hbi dataset list --app retail-ops --output json
[{ "id": "sales_daily", "metrics": 12, "dimensions": 8 }]
$ hbi dashboard create --app retail-ops "East China command center"
Created dashboard dsh_2048
$ hbi element chart create --dashboard dsh_2048 --app retail-ops --dataset sales_daily line
Created element chart_17
JSONDashboardHQLAgent-ready
Before permissions, deployment, or migration steps enter automation, both the agent and a human get a reviewable preview surface first.
$ hbi authorize grant app app_42 --user 123:editor --dry-run
Dry run passed · changes not applied
Preview: editor access -> retail-ops
--dry-runAuthorizeAudit trail
Results are not trapped in the terminal. The web interface can reflect agent changes in real time for review.
# agent updates dashboard filters from a spec file
$ hbi element filter update filter_9 --dashboard dsh_2048 --app retail-ops --file filter.yaml
SSE broadcast published
Web UI refreshed · business owner can review instantly
SSEWeb UIHuman-in-the-loop
Repo-managed skills
Give the agent both the command surface and the skills bundle for stable execution
HENGSHI CLI does not ship only a binary. It also ships the repo-managed skills directory that predefines command entry points, help discovery order, dry-run checkpoints, and cross-domain transitions. That lets both coding agents and persistent runtimes reuse the same BI runbook.
Split complexity by resource domain
Core, app, data, dashboard, data-agent, pipeline, scheduler, and workflow domains become 16 repo-managed skills instead of one overstuffed mega-prompt.
Help-first protocol
Skills begin with hbi --help and subcommand help before execution, reducing parameter guesswork.
Reusable across agent styles
Claude Code, Codex, OpenClaw, Hermes Agent, and CI pipelines can all reuse the same execution boundaries.
The CLI repo ships with 16 repo-managed skills so domain boundaries, representative commands, and execution order are already stabilized for agents.
# bundled with the CLI repo
hbi-core auth / preferences / output / terminology
hbi-data connection -> dataset -> metric / measure
hbi-dashboard plan scaffold -> validate -> apply
hbi-data-agent prompt / vector / chat config governance
hbi-pipeline node graph / schedule / integration flow
hql-expert business ask -> HE / HQL expression
hbi-workflow sequence cross-domain steps and checkpoints
agent follows domain boundaries instead of guessing commands
16 skillsHelp-firstDry-runDomain-aware
Foundation
Unify auth, app space, and governance boundaries first
hbi-core, hbi-app, hbi-permission, and hbi-user-mgmt clarify auth, app space, authorization, and org governance before the agent starts mixing app actions with platform permissions.
hbi auth statushbi app portal --app <id> show --output yamlhbi authorize grant app <id> --user 123:editor --dry-run Data & Semantics
Separate data access, semantics, and indicator domains
hbi-data, hbi-data-modeling, hbi-indicator-center, and hql-expert split connections and datasets, join design, subject-centric indicators, and HE/HQL expression authoring.
hbi dataset list --app <id>hbi data-model query --app <id> --dataset <id> "SUM({amount})"hbi subject --help Delivery
Dashboard delivery, taste, and alerts stay in their own lanes
hbi-dashboard, hbi-dashboard-taste, hbi-data-alert, and hbi-data-agent cover dashboard authoring, page composition, chart alerts, and Data Agent back-office governance.
hbi dashboard plan apply --app <app-id> --file plan.yamlhbi data-agent config --output jsonhbi data-alert --help Automation
Turn integration, notebooks, and scheduling into one chain
hbi-pipeline, hbi-notebook, hbi-scheduler, and hbi-workflow let agents connect ETL, notebook execution, existing schedule governance, and cross-domain runbooks into one reviewable flow.
hbi pipeline --helphbi notebook --helphbi scheduler --help Agent workflow
Not a chat wrapper, but an executable BI engineering interface
The value of HENGSHI CLI is that agents read context first, execute next, and hand the result back for human review. The whole chain works like an engineering interface reusable through skills and runbooks, not like scattered click-memory.
01
Read the analytics context before acting
The agent should first understand the app, dataset, semantic layer, and asset state so the next step does not become a blind mutation.
$ hbi connection list
$ hbi dataset list --app retail-ops
$ hbi data-model query --app retail-ops --dataset sales_daily "SUM({amount})"
result: 24510931
ConnectionDatasetHQL
02
Generate deliverable assets, not just answers
Once context is confirmed, the CLI stops being a chat helper and actually creates dashboards, themes, elements, and layouts that teams can ship.
$ hbi dashboard create --app retail-ops "Sales command center"
$ hbi dashboard theme show CLASSIC --as-spec --output yaml
$ hbi element container create --dashboard dsh_2048 --app retail-ops --file layout.yaml
Container + chart scaffolding ready
DashboardThemeElement
03
Keep security, permissions, and delivery in one chain
Enterprise BI engineering does not end at "created successfully". Permissions, auditability, migration, and human approval still need to be part of the same executable flow.
$ hbi authorize get app app_42
$ hbi user-group list
$ hbi authorize grant app app_42 --user 123:editor --dry-run
Ready for pipeline approval
AuthorizeUserCI-ready
Agent runtime
Built for agents that can call tools, not one specific shell
Today the obvious categories are coding agents like Claude Code and Codex, and persistent agents like OpenClaw and Hermes Agent. The more important point is future compatibility: as long as the next generation can call shell tools, HENGSHI CLI can remain their BI execution layer.
Shell-tool compatible
Fits local terminals, IDE agents, persistent services, and automation pipelines.
Skill-friendly
Works naturally with skills, template repos, runbooks, and approval workflows.
Human review built in
Dry-run previews, structured output, and SSE feedback keep automation inspectable.
Claude Code / Codex
Best for complex delivery work that stays close to a repository, shell, and spec files.
When the agent works inside a local repo or CI environment, HENGSHI CLI can turn data preparation, dashboard planning, and permission changes into one reviewable command chain.
In-repo implementation, script generation, CI/CD integration
OpenClaw / Hermes Agent
Best for persistent runtimes that live in cloud services, chat channels, or long-running automation surfaces.
These runtimes behave like an always-on execution layer: they can accept natural-language requests, keep context across sessions, and keep calling HENGSHI CLI as a stable BI tool surface.
Persistent assistants, message channels, cloud automation
Many more agents ahead
The current list is only a snapshot. Better and stronger agent runtimes will keep appearing.
The stable thing is not one shell wrapper but the shell / tool-calling interface itself. As long as the next generation of agents can call tools, HENGSHI CLI can remain their BI execution layer.
Future coding agents, persistent agents, and orchestration runtimes
Further reading
Go deeper with the existing HENGSHI CLI technical articles
HENGSHI CLI debuts as a Rust-based execution layer for Agentic BI
Why BI in the agent era should not stop at a chat shell, and why a real terminal execution layer matters.
Why HENGSHI CLI goes beyond APIs for agent-driven BI automation
A focused walkthrough of command unification, structured output, secure credential handling, and why the execution model fits agents better than API stitching.
How HENGSHI CLI supports end-to-end BI engineering delivery
A scenario-driven view of implementation, operations, permissions, and cross-environment handoff with HENGSHI CLI.
Recommended Base
HENGSHI BOX: The Ideal Physical Platform for Agent Execution
HENGSHI CLI requires a stable, efficient, and manageable runtime environment. HENGSHI BOX provides a dedicated hardware platform for Agents: pre-installed CLI environment, local LLM inference, zero data leakage—a premium base for Agent workloads.
Hardware-Level Stability
xFusion server-grade quality provides a solid foundation for long-running Agent tasks.
Local Inference
Built-in GPU/NPU acceleration enables fast local LLM response with zero public API dependency.
Plug & Play
Pre-installed HENGSHI CLI, Agent Skills suite, and vector database—just power on to dispatch Agents.
Physical Security Boundary
Data never leaves the box, network isolation, compliant with enterprise security requirements.
box-agent-run.sh
$ hbi auth status --output json
{ "status": "authenticated", "provider": "local-model" }
$ box-agent start --skill hbi-data --daemon
Agent started in background · PID: 42817
$ hbi dataset list --app retail-ops --output json
[{ "id": "sales_daily", "status": "ready" }]
$ box-agent task submit "生成华东区销售仪表盘"
Task queued · Local inference · Data stays in box
Local Model Data-in-box GPU Accelerated