Claude Code plugin

The bino-report plugin packages the bino mcp server into a batteries-included Claude Code experience: zero-config MCP registration, the IBCS knowledge the raw tools don't carry, slash commands, and a checkpointed autopilot pipeline. Because a bino report is plain-text YAML + SQL, Claude can author it directly — and bino's schema and validation are the guardrails.

Prerequisite

The plugin drives the bino binary on your PATH, so you need a build new enough to have the mcp subcommand — check with bino version, and confirm bino mcp --help works. The plugin detects a missing or too-old binary and tells you — it can't install bino for you.

Install

claude plugin marketplace add bino-bi/bino-cli
/plugin install bino

Installing registers the bino MCP server automatically — no claude mcp add step. Open /mcp to confirm the bino server is connected, and /help to see the /bino:* commands.

Mode A — Co-authoring

You drive; Claude assists. Claude follows bino's authoring loop (read the live schema → learn a dataset's columns → draft → validate_draft → write → validate_project → build) and applies IBCS semantics — scenarios (ac/pp/fc/pl), variances (d_/dr_), component choice, and data-aware narrative.

Command	What it does
/bino:report	Co-author a report end-to-end: your data + a goal → a built PDF, step by step.
/bino:new	Scaffold a fresh report bundle (init_bundle).
/bino:add-source	Probe a CSV / Excel / database source, then scaffold a typed DataSource (+ DataSet).
/bino:fix	Validate the project and walk the diagnostics to green.
/bino:build	Render the report artefacts to PDF.

Two skills load automatically when relevant: bino-authoring (the draft→validate→write discipline) and bino-ibcs (the IBCS rubric — scenarios, variances, component choice, narrative).

Mode B — Autopilot

/bino:autopilot <data-or-source> "<goal>"

A main-thread orchestrator delegates the requirements → data → authoring → validation → build pipeline to three headless subagents — bino-data, bino-author, bino-validation — pausing for you only at gates. (Only the main thread can ask you a question, so the orchestrator and requirements elicitation live there; the subagents are isolated, single-responsibility workers.)

Autonomy tiers

Tier	Routine gates	Status
0 — Supervised	confirm every step	the /bino:* co-author commands
1 — Checkpointed	pause after the brief, after the data plan, before the build	autopilot default and ceiling
2 — Autonomous	none	deferred — needs engine safety primitives that don't exist yet

The four checkpoints

Autopilot pauses for you at:

1. the report brief — after eliciting requirements (audience, message, scenarios, variance, …);
2. the data plan — after sources and datasets, flagging credentialed sources and anything the data can't satisfy;
3. before the build — rendering is slow and writes files; and
4. the finished PDF — a mandatory visual sign-off, every run.

Hard safety gates

These hold at every tier:

• Credentialed sources (databases / S3 / WebDAV) are a hard human stop. The agent writes only the *FromEnv skeleton — never an inline secret — and never spends a credential unattended.
• Running a data validation executes the agent's SQL (DuckDB SQL is not read-only), so it runs once, only on datasets authored that run, never unattended against a credentialed source.
• A non-empty / pre-existing project drops autopilot to confirmed writes — bino has no rollback yet, so it never silently clobbers hand-written manifests.
• Every build is gated and iteration-capped, and the finished PDF always needs your eyes: "validates + builds" is not the same as "correct."

A PASS from autopilot means mechanically correct and ready for sign-off — never "done."

Other MCP clients

Claude Desktop, Cursor, and other MCP clients can register bino mcp directly (see bino mcp), but they don't get the plugin's skills, commands, or autopilot — those are Claude Code only.