Overview

Blackbox AI is a multi-agent coding platform that unifies Claude Code, Codex, Gemini, and Blackbox's own models into a single workflow. Its standout architectural feature is the "Chairman" pattern — tasks are dispatched to multiple models simultaneously, and a specialized judge LLM evaluates all outputs to select the best implementation.

The Blackbox CLI is open-source (GPL-3.0, TypeScript, forked from Gemini CLI) and provides a terminal-based interface for multi-agent code generation with a built-in judge mechanism. It supports non-interactive mode, MCP, checkpointing/resume, and vision model switching — making it viable as another delegatable agent alongside our existing `claude-code` and `codex` skills.

This issue proposes a Blackbox CLI skill following the same pattern as the existing autonomous agent skills, enabling Hermes users who subscribe to Blackbox AI to delegate coding tasks through their multi-model pipeline.

---

Research Findings

How the Blackbox CLI Works

Architecture: The CLI is a Node.js (v20+) application built on the Google Gemini CLI codebase (`gemini.tsx` core), extended with:

• SubagentManager (`packages/core/src/subagents/`) — Orchestrates multiple sub-agents with events, hooks, statistics tracking, and validation
• Built-in agents (`builtin-agents.ts`) — Pre-configured agent definitions for Claude, Codex, Gemini, and Blackbox models
• Non-interactive mode (`nonInteractiveCli.ts`) — Full headless execution with tool calls, checkpointing, and structured output formatting. Separates thoughts (stderr) from results (stdout)
• Judge mechanism — Evaluates competing agent outputs and selects the best implementation
• Checkpoint system — Save/resume sessions via `.blackboxcli/checkpoint-*.json` files
• MCP support — Model Context Protocol integration for external tool connectivity

Installation:
```bash

From npm (prebuilt)

npm install -g @blackboxai/cli

From source

git clone https://github.com/blackboxaicode/cli.git
cd cli && npm install && npm install -g .

Configure

blackbox configure # Enter API key from app.blackbox.ai/dashboard
```

Non-interactive usage (key for Hermes delegation):
```bash

One-shot task execution

blackbox --prompt "Implement JWT authentication for the Express API"

Resume from checkpoint

blackbox --resume-checkpoint "task-abc123" --prompt "Add refresh token support"
```

Session commands: `/compress` (shrink history), `/clear` (reset), `/stats` (token usage)

Key Design Decisions

1. Gemini CLI as base — Rather than building from scratch, Blackbox forked Google's Gemini CLI for its tool execution engine, event handling, and terminal UI (React/Ink). This gave them a production-quality foundation.
2. Multi-model by default — The CLI can run the same prompt through multiple models (Blackbox Pro, Claude Sonnet 4.5, GPT-5.2 Codex, Gemini 2.5 Pro) and use a judge to pick the best.
3. Credit-based pricing — Free tier has basic access; Pro ($10/mo) adds $30 credits for premium models; Pro Plus ($20/mo) unlocks multi-agent execution.
4. GPL-3.0 license — The CLI is copyleft, which is important to note for any integration approach.

Platform Capabilities Beyond the CLI

The Blackbox platform also offers:

• Multi-Agent Task API (`POST cloud.blackbox.ai/api/tasks`) — Programmatic dispatch to 2-5 agents with automated judge selection. Requires `bb_xxxxxxxx` API key.
• Semantic Knowledge Graph — Indexes entire repos, commits, and URLs for cross-file reasoning
• Isolated sandboxes — Long-running tasks in background environments
• IDE support — VS Code extension with 4.2M+ installs, JetBrains, Blackbox IDE

---

Current State in Hermes Agent

Existing agent delegation skills:

Skill	Agent	Installation	Non-interactive	Auth Required
`claude-code`	Claude Code CLI	`npm install -g @anthropic-ai/claude-code`	Yes (`--dangerously-skip-permissions`)	Anthropic API key
`codex`	OpenAI Codex CLI	`npm install -g @openai/codex`	Yes (`exec` subcommand)	OpenAI API key
`hermes-agent`	Hermes Agent	Already installed	Yes (`-q` flag)	Configured LLM provider
`blackbox` (proposed)	Blackbox CLI	`npm install -g @blackboxai/cli`	Yes (`--prompt`)	Blackbox API key

No existing Blackbox integration. Zero mentions of "blackbox" in the hermes-agent codebase.

Related issues:

• Feature: Cross-CLI Agent Orchestration — Mixed Workflows with External Agent CLIs (inspired by AgentWorkforce/relay) #413 — Cross-CLI Agent Orchestration (framework for mixed CLI agent workflows — Blackbox CLI would be another backend)
• Feature: Multi-Agent Architecture — Orchestration, Cooperation, Specialized Roles & Resilient Workflows #344 — Multi-Agent Architecture (orchestration infrastructure)
• Feature: Consensus & Voting Engine for Multi-Agent Decision Making (inspired by AgentWorkforce/relay) #412 — Consensus & Voting Engine (related but distinct from Blackbox's judge pattern — see sister issue)

---

Implementation Plan

Skill vs. Tool Classification

This should be a skill because:

• It wraps an external CLI (`blackbox`) callable via `terminal(pty=true)`
• Follows the exact same pattern as existing `claude-code` and `codex` skills
• No custom Python integration needed — all interaction is through the terminal
• API key management is handled by the user running `blackbox configure`

Not bundled — Skills Hub. Blackbox AI requires a paid subscription, has GPL-3.0 licensing, and modest adoption (191 stars). It's specialized enough to be a Skills Hub skill rather than bundled.

What We'd Need

1. Skill file (`SKILL.md`) following the `claude-code`/`codex` pattern
2. Prerequisites section — Node.js 20+, npm, Blackbox API key
3. One-shot mode — `blackbox --prompt "task"` via `terminal(pty=true)`
4. Background mode — Long-running sessions with `process(poll/log)` monitoring
5. Checkpoint/resume — Leverage Blackbox's built-in checkpoint system
6. Multi-model hints — Document how to configure which models run (agent selection)
7. PR review pattern — Adapt the existing PR review workflow to Blackbox's capabilities

Phased Rollout

Phase 1: Basic Delegation Skill

• SKILL.md with one-shot and background delegation patterns
• Prerequisites, configuration, and usage documentation
• Standard PR review workflow adaptation
• Metadata: tags, related_skills (claude-code, codex, hermes-agent)

Phase 2: Multi-Model Features

• Document Blackbox's built-in multi-agent/judge mode for skill users
• Add guidance for when to use Blackbox's multi-model mode vs. Hermes's native delegation
• Checkpoint/resume workflow for long-running tasks

Phase 3: Cross-CLI Integration

• When Feature: Cross-CLI Agent Orchestration — Mixed Workflows with External Agent CLIs (inspired by AgentWorkforce/relay) #413 lands, register Blackbox as an `ExternalCLIBackend` in the cross-CLI orchestration framework
• Enable mixed workflows: Blackbox + Claude Code + Codex in a single workflow DAG

---

Pros & Cons

Pros

• Consistent pattern — Follows the established claude-code/codex skill architecture exactly
• Multi-model value — Users get access to Blackbox's built-in judge/multi-model pipeline, which is unique among CLI agents
• Low effort — ~200 lines of SKILL.md, no code changes needed
• MCP compatibility — Blackbox CLI speaks MCP, potentially enabling tool sharing with Hermes
• Checkpoint/resume — Built-in session persistence for long-running tasks

Cons / Risks

• GPL-3.0 license — Copyleft license on the CLI; doesn't affect a skill (we're just calling it), but worth noting
• Paid service — Requires Blackbox subscription ($10-40/mo) with credit-based consumption
• Modest adoption — 191 GitHub stars, forked from Gemini CLI; uncertain long-term maintenance
• Wrapper-of-wrappers concern — Blackbox CLI itself wraps Claude Code, Codex, and Gemini. If Hermes already delegates to those directly, the added value is mainly the judge layer
• Last commit Dec 2025 — ~3 months since last meaningful update; maintenance risk
• TrustPilot 1.9/5 — Mixed user reviews, primarily around billing transparency and credit consumption

---

Open Questions

• Is the Blackbox CLI's built-in judge mechanism valuable enough to justify using it over direct claude-code/codex delegation + a native Hermes judge (see related best-of-N issue)?
• Should we wait for Feature: Cross-CLI Agent Orchestration — Mixed Workflows with External Agent CLIs (inspired by AgentWorkforce/relay) #413 (Cross-CLI Orchestration) before creating individual agent skills, or proceed independently?
• Does the GPL-3.0 license on the CLI create any concerns for Skills Hub distribution?

---

References

• Blackbox AI Website
• Blackbox CLI GitHub (GPL-3.0) — 191 stars, TypeScript, forked from Gemini CLI
• Blackbox CLI Docs
• Multi-Agent Execution Docs
• Multi-Agent Task API
• Blackbox AI Review 2026 (Banani) — Detailed independent review
• Existing Hermes skills: `claude-code`, `codex`, `hermes-agent` (autonomous-ai-agents category)
• Related issues: Feature: Cross-CLI Agent Orchestration — Mixed Workflows with External Agent CLIs (inspired by AgentWorkforce/relay) #413, Feature: Multi-Agent Architecture — Orchestration, Cooperation, Specialized Roles & Resilient Workflows #344, Feature: Consensus & Voting Engine for Multi-Agent Decision Making (inspired by AgentWorkforce/relay) #412
  ISSUE_EOF; __hermes_rc=$?; printf 'HERMES_FENCE_a9f7b3'; exit $__hermes_rc