Import Bundle
Bring a whole Claude Code plugin or skill repository into Hrida AI Studio in one pass.
Import Bundle reads a .zip export of a Claude Code plugin marketplace or skill collection — SKILL.md files, agent definitions, slash commands, agent-orchestrator cookbooks, reference docs, bundled scripts — and maps every recognized file to the matching Hrida entity automatically, based on its folder structure. You review exactly what was detected before anything is created, and each item can be checked or unchecked individually.
This is the fast path for migrating an existing plugin repo instead of recreating dozens of skills, tools, and prompts by hand.
Two modes
Import Bundle behaves differently depending on where you open it from — the same .zip file, scoped to different target entities:
| Mode | Entry point | Creates |
|---|---|---|
| Workspace mode | Workspace → Skills / Tools / Prompts, via the Import Bundle (.zip) button | Workspace Skills, Tools, and Prompts — no Agent Skills, no Space, no Workflows or Knowledge |
| Agent Builder mode | Agent Builder drafts list or Agent Builder → Skills, via the Import Bundle (.zip) button | Agent Skills, Prompts, Agent Workflows (one per orchestrator cookbook), and optional reference-doc Knowledge bases |
Both modes are reached the same way — drop or choose a .zip file — and share the same review-then-confirm flow. The only difference is which bucket of entity types the zip's contents get classified into.
How it works
1. Drop the zip
Drop a .zip file onto the page, or click Choose File. The zip is parsed entirely in the browser (via JSZip) — nothing is uploaded until you confirm the import in step 3.
2. Review
Every recognized file is classified and grouped into sections. Each item has a checkbox — uncheck anything you don't want imported. Sections only appear if the bundle contains that type of content:
| Section | Detected from | Notes |
|---|---|---|
| Skills | SKILL.md files | Workspace mode only |
| Tools | scripts/*.py bundled alongside a skill | Workspace mode only — see code execution warning below |
| Agent Skills | SKILL.md and agents/<name>.md files | Agent Builder mode only; lists any source tool names that don't map to a Hrida built-in tool (dropped, not silently invented) |
| Prompts | commands/*.md slash-command files | Both modes |
| Agent Workflows | agent.yaml + subagents/*.yaml cookbooks | Agent Builder mode only — one workflow per cookbook, built as a linear chain (orchestrator → each subagent in sequence); flagged as an approximation to rewire in the graph editor, not a claim of correct control flow |
| Knowledge (opt-in) | references/*.md docs bundled with a skill | Agent Builder mode only; unchecked by default — check the groups you actually want uploaded |
| Other skill documents | Loose .md files with no recognizable frontmatter | Falls back to a plain workspace Skill |
The review screen also shows:
- MCP servers referenced — listed for information only; never auto-registered, since they need real credentials the zip can't supply. Configure them manually under Settings → Tool Servers if you have credentials.
- Malformed
.mcp.jsonwarning — if an.mcp.jsonfile can't be parsed, it's skipped with a note; the rest of the import is unaffected. - Re-import duplication notice — re-importing the same zip later creates duplicate skills/tools/prompts. Only workflow-embedded skills are deduplicated (by name, server-side, same as workflow import).
3. Execute
Click Import Selected. Each checked item is created independently, one API call at a time, in its own try/catch — one failing item never blocks the rest of the batch. The final screen lists every item with a ✓ or ✗ and an error detail for anything that failed.
Tools run real code
Each detected tool wraps its bundled Python script and runs it as-is via subprocess — behavior is unchanged from the original .py file. Only import a bundle if you fully trust its source. This is the same warning shown for any manually created Tool, surfaced here per-item before you confirm.
Tool name translation (Agent Skills)
Claude Code tool names (Read, Grep, Bash, mcp__*, …) don't map one-to-one onto Hrida's closed Agent Skill tool set. Only two have honest equivalents:
| Source tool | Mapped to |
|---|---|
WebSearch | web_search |
Bash (when the skill bundles a scripts/ folder) | code_executor |
Every other source tool name (Read, Grep, Glob, Edit, Write, mcp__*) is dropped, and every drop is listed per-item on the review screen under "tools not representable in Hrida" — never silently invented and never silently discarded without a trace.
What is not imported
| Content | Reason |
|---|---|
claude-for-msft-365-install/, scripts/, .github/, .githooks/ | Repo tooling, not plugin content |
CLAUDE.md, README.md, LICENSE | Repo documentation |
.mcp.json server definitions | Informational only — see above |
.claude-plugin/plugin.json, marketplace.json, steering-examples.json | Metadata only — used to enrich names/tags, not persisted as their own entity |
Use Cases
Migrating a Claude Code plugin marketplace
Point Import Bundle at an exported plugin repo .zip and get every skill, slash command, and agent definition mapped into Hrida in one review-and-confirm pass instead of recreating each one by hand.
Bringing in a script library as Tools
A folder of standalone .py scripts bundled under a skill's scripts/ directory becomes a set of workspace Tools automatically, each one just shelling out to the original script — see Tools.
Prototyping an orchestrator workflow from a cookbook
An agent.yaml + subagents/*.yaml cookbook becomes a draft Agent Workflow with the right skills embedded and wired in sequence — a starting point to rewire in the graph editor rather than a finished production workflow.
Limitations
No dedup outside of workflows
Standalone Skills, Agent Skills, and Prompts created by an import have no dedup against a previous import of the same zip — re-running the same bundle creates duplicates. Only skills embedded in an Agent Workflow are deduplicated (by exact name match), matching the workflow import behavior it reuses.
Linear-chain workflow approximation
Cookbook workflows are built as a straight chain (start → orchestrator → subagent-1 → … → end). This does not attempt to reconstruct conditional branching, parallelism, or looping that the original cookbook's orchestrator logic may have had — treat the imported graph as a starting point.
Knowledge base content, not the base itself, requires opt-in
Reference-doc groups are unchecked by default specifically because they create and upload content immediately on confirm, unlike the other sections which just create metadata-only entities.