Feature: Parallel CLI Skill — Web Search, Extraction, Deep Research & Enrichment via parallel-cli

[kshitijk4poor]

Overview

Parallel just released a standalone parallel-cli that gives terminal-native agents direct access to Parallel's web intelligence stack: web search, clean content extraction, deep research, and data enrichment from the command line.

This is a strong fit for Hermes because we already have a recurring pattern of wrapping high-value external CLIs in bundled skills rather than hard-coding one-off integrations. parallel-cli is especially attractive because it is explicitly designed for agent use: structured --json output, non-interactive execution, async --no-wait flows for long-running tasks, and first-party skill definitions in the parallel-agent-skills repo.

The proposal is to add a bundled Hermes skill for parallel-cli, following the same philosophy as mcporter or codex: teach Hermes when to install it, how to authenticate, which subcommands map to which user intents, and how to use the JSON outputs safely.

This is directly implementable today using existing Hermes capabilities (terminal, process, file tools) and requires zero core code changes — only a new skill.

---

Research Findings

What Parallel CLI Provides

From the launch post and repo, parallel-cli exposes:

• search — web search with structured JSON output
• extract / fetch — clean markdown extraction from arbitrary URLs
• research run|status|poll|processors — long-running deep research jobs with async polling
• enrich run|status|poll|plan|suggest|deploy — AI-driven data enrichment workflows
• findall — web-scale entity discovery
• monitor — continuous web change tracking
• login, logout, auth — auth lifecycle for interactive or headless use

The standalone binary supports installation via:

brew install parallel-web/tap/parallel-cli
curl -fsSL https://parallel.ai/install.sh | bash
npm install -g parallel-web-cli

And the Python package supports:

pip install "parallel-web-tools[cli]"

Why It Fits Hermes Well

Parallel explicitly designed the CLI for terminal-based agents:

• Every major command supports --json
• Long-running operations support async launch + status polling
• Commands work non-interactively
• Input can be piped via stdin
• The vendor already ships agent-oriented skill definitions in parallel-web/parallel-agent-skills

That means Hermes does not need to invent usage patterns from scratch — we can adapt the upstream patterns into our own skill format.

Upstream Skill Patterns We Can Borrow

Parallel's own agent skills split the workflow conceptually into:

• setup/auth
• web search
• URL/content extraction
• deep research
• task status / result retrieval
• enrichment

Their search skill defaults to JSON output, saves artifacts for follow-up work, and requires inline citations sourced from the CLI output. That maps cleanly onto Hermes skill guidance.

---

Current State in Hermes Agent

What we have:

• Native web_search and web_extract tools for general-purpose web retrieval
• terminal + process tools for driving third-party CLIs
• Existing skill patterns for external CLIs (mcporter, codex, claude-code)
• Strong support for JSON-first command-line workflows

Gap:

• No Hermes skill for parallel-cli
• No documented workflow for using Parallel's search/extract/research/enrichment stack from Hermes
• No built-in guidance on install/auth, async polling, or when to prefer Parallel over native web tools

This means users currently have to discover and operate parallel-cli manually, despite it being a natural fit for Hermes's terminal-driven workflow model.

---

Implementation Plan

Skill vs. Tool Classification

This should be a bundled skill, not a core tool, because:

• The capability is already exposed through a mature CLI
• Hermes can access it through existing terminal and process tools
• No new Python integration, schema plumbing, or API key loader is required
• The workflow is primarily procedural instruction: install, authenticate, call subcommands, parse JSON, poll async jobs
• It is broad enough to be useful, but not foundational enough to justify new core tool surface area yet

Proposed Scope

Add a new bundled skill, likely at:

skills/research/parallel-cli/SKILL.md

The skill should cover:

1. Installation and auth

   • install via brew / install script / pip fallback
   • login interactively or use PARALLEL_API_KEY
   • verify with parallel-cli auth --json

2. Search workflow

   • parallel-cli search ... --json
   • when to use it vs Hermes web_search
   • citation discipline from returned URLs only

3. Extraction workflow

   • parallel-cli extract URL --json
   • using --objective and --full-content

4. Deep research workflow

   • parallel-cli research run ... --json
   • --no-wait + status / poll
   • when to prefer async task flow

5. Enrichment / FindAll / Monitor overview

   • enough guidance to use the CLI safely for structured workflows
   • JSON-first examples Hermes can copy directly

6. Rules / pitfalls

   • always prefer --json
   • prefer non-interactive flags for agent usage
   • use background/process tools only for genuinely long-running flows
   • never cite URLs not returned by the CLI output

Phased Rollout

Phase 1: Single comprehensive skill

• Add parallel-cli skill with install/auth/search/extract/research/enrichment guidance
• Focus on the highest-value terminal workflows

Phase 2: Optional refinement

• Split into multiple more specialized skills if real usage warrants it (parallel-search, parallel-research, etc.)
• Add trigger guidance for when Parallel should be preferred over native web tools
• Add artifact-handling conventions for saved JSON outputs

---

Pros & Cons

Pros

• Zero core changes required
• Adds immediate support for a newly launched agent-native CLI
• Reuses an existing Hermes pattern for external CLIs
• Unlocks richer web workflows than the current basic search/extract pair
• Lets Hermes users benefit from Parallel's deep research, enrichment, and monitoring primitives without bespoke integration work

Cons / Risks

• Requires the user to install/auth a third-party CLI
• Overlap with existing Hermes native web tools may create ambiguity about which path to choose
• A single broad skill may become long if it tries to cover every subcommand in depth
• Some workflows (research, monitoring, enrichment) are long-running and need careful async guidance to avoid poor UX

---

Open Questions

• Should the first version be a single parallel-cli skill, or multiple narrower skills mirroring Parallel's own upstream split?
• Should Hermes prefer this skill only when the user explicitly mentions Parallel, or proactively when a task needs deeper research/enrichment than native web tools provide?
• Should we document findall and monitor fully in v1, or keep the MVP focused on search / extract / research?

---

References

• Launch post: https://parallel.ai/blog/parallel-cli
• CLI repo: https://github.com/parallel-web/parallel-web-tools
• Agent skills repo: https://github.com/parallel-web/parallel-agent-skills
• Docs: https://docs.parallel.ai/home
• Related Hermes skills: mcporter, codex, claude-code

[kshitijk4poor]

Opened PR #985 implementing the new parallel-cli skill.