How to Choose Pi, Hermes, and Other Agents

Compare Pi, Hermes Agent, Claude Code, Cursor, OpenCode, Codex, and Prime Agent. Pick the right tool for a repo, a research job, or Gab AI media tools.

The Harness vs the Model

The agent tool is the harness: it owns the loop, the file tools, the terminal, and the UI. Gab AI is the model provider behind it. Run Arya, GPT, Claude, Gemini, and the rest through all of these CLIs with one Gab AI API key — not an OpenAI key, not an Anthropic key. Do this first, then pick a harness.

Run Them on the Gab AI API

Pi, Hermes, Cursor, Claude Code, Codex, OpenCode, and Prime Agent are clients. They should call https://gab.ai/v1 so one Plus subscription and one credit balance cover the whole loop, including auto routing. If the CLI is still talking to api.openai.com or api.anthropic.com, you are not using Gab AI yet.

Pi — official Gab AI package

Hermes — custom Gab AI endpoint

Do not paste an OpenAI or Anthropic key into these tools

One Gab AI key. Base URL https://gab.ai/v1 (Claude Code: https://gab.ai). Credits come off the same balance as chat, images, and video. Per-tool screens and config files are in the setup guides at the bottom of this lesson.

Then Pick a Harness

Do not start by collecting every agent. Start with the job. The list below is a field guide, not a ranking — every option below should already be wired to Gab AI.

A 30-Second Decision

If you are unsure, use this and ignore the rest of the catalogue until you have run one real job.

What Does Not Matter (Yet)

Do not pick a harness because of a benchmark screenshot. The quality of the loop is mostly the brief, the Gab AI model, and whether you review the diff. A mediocre CLI with a sharp brief beats a famous CLI with "fix the app." Switch later if the tool fights you. The Gab AI key and the brief travel with you.

Setup guides — Gab AI API

  1. Create a Gab AI API key — You need Gab AI Plus. In the app go to Settings → API and create a key. That key is the only credential the agent should see. The API keys guide is in the list at the bottom of this lesson.
  2. Point the agent at https://gab.ai/v1 — Most CLIs take a custom OpenAI-compatible base URL. Set it to https://gab.ai/v1 and paste the Gab AI key. Claude Code is the exception: it uses the Anthropic-compatible Messages API with base URL https://gab.ai (no /v1 suffix).
  3. For Pi, install the official Gab package — Pi is first-class on Gab AI. Install @gabai/pi-gab-ai, then run /gab inside Pi and paste the key once. That registers Gab models and Gab API tools — images, video, speech, files — so the agent can call them mid-loop.
  4. Pick a function-calling Gab model, or auto — The loop dies without tool support. Use auto, or a slug from GET https://gab.ai/v1/models with function_calling. Display names like "Claude Opus 4.8" return 400; use the hyphenated slug.
pi install npm:@gabai/pi-gab-ai
pi --models "gab/*"
# Inside Pi: /gab   (paste your Gab AI API key once)
hermes model
# Custom endpoint (self-hosted / VLLM / etc.)
#   API base URL: https://gab.ai/v1
#   API key:      YOUR_GAB_API_KEY
#   Model name:   auto   (or arya, claude-sonnet-4-5, …)

Do not paste an OpenAI or Anthropic key into these tools

One Gab AI key. Base URL https://gab.ai/v1 (Claude Code: https://gab.ai). Credits come off the same balance as chat, images, and video. Per-tool screens and config files are in the setup guides at the bottom of this lesson.

  1. Stay in the browser → Gab custom agents — No terminal and no API key. A named specialist with a system instruction, optional pinned model, and your memory. Best for writing, support replies, and any job you repeat in chat. Covered in Custom Agents.
  2. Want Gab tools inside the loop → Pi — Pi is a lightweight coding CLI. The official @gabai/pi-gab-ai package registers Gab models and Gab API tools — image generation, video, speech, files, credits, embeddings — so the agent can call them mid-loop instead of you bouncing back to the website.
  3. Want a general agent, not just a coder → Hermes — Hermes Agent (Nous Research) is a general-purpose agent CLI. Point it at https://gab.ai/v1 with your Gab AI key, pick a function-calling model, and it will use tools across research, files, and code. Use it when the job is not "edit this repo" so much as "go figure this out and come back with artifacts."
  4. You live in the editor → Cursor — Cursor is an IDE. The agent can see the buffer you are in, the git diff, and the linter. Add Gab AI as a custom OpenAI-compatible provider (https://gab.ai/v1 plus your Gab AI key) when you want Arya or Auto routing without leaving the file. Best when you want to steer line by line.
  5. Repo work in the terminal → Claude Code, Codex, OpenCode — These are coding-first CLIs. Claude Code speaks Gab AI's Anthropic-compatible Messages API. Codex uses the OpenAI Responses API on https://gab.ai/v1. OpenCode uses Chat Completions on the same base URL. Pick the one whose workflow you already know — the Gab-side difference is the base URL and the Gab AI key, not the prompting style.
  6. Long-running research or coding → Prime Agent — Prime Agent (Prime Intellect) is built for jobs that should keep going: a research harness, a large refactor, a task you would otherwise babysit. Wire Gab AI as a custom OpenAI-compatible provider (https://gab.ai/v1) and give it a scoped goal with a hard stop condition.