Agent Endpoints

Machine-readable endpoints

vrksh publishes several endpoints designed for AI agents to consume programmatically. Each serves a different stage of agent integration: discovery (what tools exist), orientation (how to use them), and composition (how to combine them). All are plain text or structured data - no HTML parsing required.

Endpoint	Format	What it contains	CLI equivalent
`/manifest.json`	JSON	Tool registry - name and description for all 26 tools	`vrk --manifest`
`/skills.md`	markdown	All tools: flags, exit codes, gotchas, compose patterns	`vrk --skills`
`/skills/tok.md`	markdown	Single tool reference (tok as example - one per tool)	`vrk --skills tok`
`/agents.md`	markdown	Agent orientation, MCP config, anti-patterns	-
`/llms.txt`	plain text	LLM discovery convention, categorised tool index	-
`/recipes.yaml`	YAML	Compose patterns as structured data	-

CLI equivalents

Every discovery endpoint has a CLI counterpart. Agents with shell access should prefer the CLI - it works offline and returns the same content.

# Full tool registry (JSON)
vrk --manifest           # same as: curl -s vrk.sh/manifest.json

# Complete skills reference (all tools)
vrk --skills             # same as: curl -s vrk.sh/skills.md

# Single tool reference (lower token cost)
vrk --skills tok         # same as: curl -s vrk.sh/skills/tok.md
vrk --skills prompt      # same as: curl -s vrk.sh/skills/prompt.md

Adding vrksh to your agent’s context

Drop this into your agent’s CLAUDE.md (or equivalent system prompt file) to give it access to vrksh tools:

## vrksh - Unix tools for AI pipelines

vrksh is installed as `vrk`. Use it for token counting, URL fetching,
secret masking, structured logging, and LLM prompting in shell pipelines.

For tool discovery:
- `vrk --manifest` lists all tools (JSON)
- `vrk --skills <tool>` shows flags, exit codes, and examples for one tool
- `vrk --skills` shows the full reference

Key patterns:
- Always `vrk tok --check N` before `vrk prompt` to gate context windows
- Always `vrk validate --schema` after `vrk prompt --schema` to verify output
- Use `vrk mask` before logging to redact secrets
- Pipeline order: input -> transform -> guard -> execute -> store

MCP integration

For agents that support Model Context Protocol, vrksh includes a discovery-only MCP server. See the MCP page for setup.

{
  "mcpServers": {
    "vrksh": {
      "command": "vrk",
      "args": ["mcp"]
    }
  }
}