Skip to main content

​
Project Setup

The tank init command creates a new skills.json manifest for publishing AI agent skills.

​
tank init

Interactively create a skills.json manifest in the current directory. 
tank init

​
Interactive Prompts

$ tank init
? Skill name: my-skill
? Version: 0.1.0
? Description: AI skill for data processing
? Make this skill private? No
? Author: Alice Developer
✓ Created skills.json

​
Generated Manifest

{
  "name": "my-skill",
  "version": "0.1.0",
  "description": "AI skill for data processing",
  "visibility": "public",
  "skills": {},
  "permissions": {
    "network": {
      "outbound": []
    },
    "filesystem": {
      "read": [],
      "write": []
    },
    "subprocess": false
  }
}

​
Field Validation

​
Name

  • Pattern: Lowercase alphanumeric + hyphens, optionally scoped
  • Examples: my-skill, @myorg/my-skill
  • Max length: 214 characters
  • Invalid: MySkill, my_skill, my skill
# Valid names
my-skill
data-processor
@acme/analytics
@org-name/skill-name-123

# Invalid names
My-Skill           # Uppercase
my_skill           # Underscores
my skill           # Spaces
-my-skill          # Leading hyphen

​
Version

  • Pattern: Semantic versioning (semver)
  • Format: MAJOR.MINOR.PATCH[-prerelease][+build]
  • Examples: 1.0.0, 2.1.3-beta, 0.1.0+20240101
# Valid versions
0.1.0
1.0.0
2.3.4-alpha
1.0.0-beta.1
2.0.0+build.123

# Invalid versions
1.0              # Missing patch
v1.0.0           # Prefix
1.0.0.0          # Too many parts

​
Scoped Packages

Use scoped names for organization-owned skills: 
$ tank init
? Skill name: @myorg/analytics
? Version: 0.1.0
? Description: Analytics skill for MyOrg
? Make this skill private? Yes  # Scoped packages default to private
? Author: MyOrg Team
✓ Created skills.json
Scoped packages:
  • Default to private visibility
  • Require organization membership to publish
  • Use format @org/name

​
Visibility

​
Public Skills

  • Visible to all users in the registry
  • Anyone can install without authentication
  • Indexed in search results
  • Default for non-scoped packages

​
Private Skills

  • Only visible to authorized users
  • Require authentication to install
  • Not indexed in public search
  • Default for scoped packages (@org/name)

​
Overwriting Existing Files

If skills.json exists, tank init prompts for confirmation: 
$ tank init
⚠ skills.json already exists in this directory.
? Overwrite existing skills.json? (y/N)
  • Yes - Replaces existing file (data is lost)
  • No - Aborts without changes

​
Default Values

  • Name: Current directory name
  • Version: 0.1.0
  • Description: Empty string
  • Visibility: public (or private if scoped)
  • Author: Name from ~/.tank/config.json (if logged in)

​
Permissions Model

The generated manifest includes an empty permissions budget: 
{
  "permissions": {
    "network": {
      "outbound": []  // Allowed domains
    },
    "filesystem": {
      "read": [],     // Allowed read paths
      "write": []     // Allowed write paths
    },
    "subprocess": false  // Subprocess execution
  }
}
See Permissions for configuration details.

​
Next Steps

After creating skills.json:
  1. Add skill files - Create your skill implementation
  2. Configure permissions - Define required permissions in skills.json
  3. Test locally - Use tank link to test with agents
  4. Publish - Run tank publish to upload to registry
# Example workflow
tank init
# ... add skill files ...
tank publish --dry-run  # Preview
tank publish            # Upload

​
Validation

tank init validates the generated manifest against the skills.json schema before writing. If validation fails: 
✗ Generated skills.json is invalid:
  name: Name must be lowercase, alphanumeric + hyphens
  version: Version must be valid semver (e.g. 1.0.0)
Common validation errors:
  • Invalid name format (uppercase, special characters)
  • Invalid version format (not semver)
  • Missing required fields

​
Manual Editing

After tank init, you can manually edit skills.json to:
  • Add dependencies to skills object
  • Configure permission budget
  • Add metadata (homepage, repository, keywords)
  • Set audit score threshold (audit.min_score)
See skills.json Reference for full schema.

Build docs developers (and LLMs) love

Get started for freeTalk to us