Free Antigravity Models in Hermes

thumbnail

date

Aug 10, 2026

slug

hermes-antigravity-blog

status

Published

Integrating Google Antigravity Models into Hermes Agent: The Complete Guide & Engineering Story

How to get free access to Gemini 3.1 Pro, Gemini 3.5 Flash, Gemini 3.6 Flash, and Claude 4.6 inside Hermes Agent—and how to fix the hidden schema, identity filter, and provider routing bugs.

💡 TL;DR: Google’s Antigravity (Cloud Code Assist) offers generous free tiers for frontier models (Gemini 3.x, Claude 4.6, GPT-OSS 120B). Integrating these models into Hermes Agent via the antigravity-provider plugin requires solving four distinct engineering challenges:
Schema Translation: Sanitizing standard JSON Schema keywords (const, oneOf, $ref) for Gemini's strict functionDeclarations.
Identity Scrubbing: Neutralizing a secret Google upstream prompt classifier that returns a masked 429 RESOURCE_EXHAUSTED whenever system instructions contain "You are Hermes Agent...".
Provider Resolution: Bridging Hermes’ runtime plugin registry with resolve_provider_full() to eliminate Unknown provider 'antigravity' errors during model switching.
Wire Model Mapping: Dynamically mapping reasoning effort levels (low, medium, high) to Google's expected wire model IDs (e.g. gemini-3.6-flash-high).
This guide covers the complete setup, architectural breakdown, and step-by-step fixes to get Antigravity running seamlessly in Hermes.

1. Introduction: Why Google Antigravity + Hermes Agent?

If you use tools like omp (Oh My Pi), you are likely familiar with Google Antigravity: free, high-throughput access to frontier AI models including Gemini 3.1 Pro, Gemini 3.5 Flash, Gemini 3.6 Flash, Claude 4.6 (via Google), and GPT-OSS 120B.

For developers who use Hermes Agent (by Nous Research) as their primary terminal coding assistant, bringing Antigravity into Hermes is a huge win—it unlocks powerful, zero-cost reasoning and coding models directly inside Hermes’ rich tool-calling and memory environment.

Hermes supports custom model providers via plugins. The antigravity-provider plugin serves as an in-process bridge, translating OpenAI-formatted chat completions into Google’s Cloud Code Assist API (v1internal:streamGenerateContent).

However, bringing these two systems together out of the box leads to immediate failures. Here is the full guide to getting it working.

2. Quickstart & Installation Setup

Getting the plugin running requires three setup steps:

Step 1: Install the Plugin in Editable Mode

Because the plugin repository (~/.hermes/plugins/antigravity-provider) uses a Python src/ layout (src/antigravity_provider), Hermes' default launcher shim may fail to locate the package:

This binds the plugin directly into Hermes' venv, ensuring imports persist across authentication refreshes.

Step 2: Authenticate OAuth Credentials

Authenticate your Google account using hermes agy login:

This retrieves your OAuth tokens and stores them securely in the macOS Keychain or ~/.hermes/.

Step 3: Configure `config.yaml`

Add the antigravity provider block to ~/.hermes/config.yaml:

3. Engineering Deep Dive: Overcoming the Integration Hurdles

Connecting Hermes to Antigravity isn't just a matter of passing credentials—it requires adapting Hermes' OpenAI-style structures to Google's specialized Cloud Code Assist backend.

Challenge 1: Translating Tool JSON Schemas

When Hermes executes an agent session, it exports JSON Schemas for its active tools. Standard Hermes schemas use modern JSON Schema features (const, oneOf, anyOf, $ref, definitions, additionalProperties).

Google’s Gemini API expects minimal OpenAPI / Gemini schema objects. Sending standard Hermes schemas results in HTTP 400 validation rejections:

Solution: Update _schema() in transform.py to recursively strip non-compliant schema keywords before translating tools upstream:

Challenge 2: Unmasking the Identity 429 Filter

With schemas sanitized, requests reached Google’s API... but every full agent turn immediately failed with a 429 RESOURCE_EXHAUSTED error:

A standard HTTP 429 implies rate limits or zero quota. However, testing the exact same OAuth credentials inside omp succeeded instantly.

I extracted the OAuth tokens and tested requests using Hermes' request builder:

Minimal Request ("Reply with: OK"): HTTP 200 OK

Full Agent Request (19KB system prompt + 36 tool definitions): HTTP 429

Small requests worked, so quota wasn't exhausted. Furthermore, sending a dummy 30KB system prompt with all 36 tools also returned HTTP 200 OK, proving total payload size was not the trigger.

The failure was content-dependent.

By systematically binary-searching Hermes’ 19,000-character system prompt line by line against the live API, I isolated the exact sentence causing the 429 rejection:

Testing variations of this sentence confirmed the behavior:

System Instruction Variation	API Result
`"You are Hermes Agent."`	✅ 200 OK
`"You are an intelligent AI assistant created by Nous Research."`	✅ 200 OK
`"You run on Hermes Agent (by Nous Research)."`	✅ 200 OK
`"You are Hermes Agent, an intelligent AI assistant created by Nous Research"`	❌ 429 EXHAUSTED
`"You are Hermes Agent, an intelligent AI assistant created by X Corp"`	✅ 200 OK
`"You are a Helpful Agent, an intelligent AI assistant created by Nous Research"`	✅ 200 OK

Google’s Cloud Code Assist API runs an upstream classifier that filters out system instructions asserting a competing assistant identity. When triggered, the API returns a masked 429 RESOURCE_EXHAUSTED status code rather than a standard policy error.

omp avoids this because it injects Google's own identity line ("You are Antigravity, a powerful agentic AI coding assistant designed by the Google Deepmind team...").

To fix this in Hermes without altering behavioral instructions, I added a targeted regex scrubber in transform.py:

Applying this scrub restored immediate HTTP 200 OK responses while preserving all of Hermes' tool guidance, formatting rules, and session memory.

Challenge 3: Solving Hermes' Dual Provider Resolution System

With single-shot requests working (hermes -z), running hermes model or switching models in the TUI (/model) raised:

Hermes maintains two distinct provider resolution paths:

resolve_runtime_provider(): Used for execution. It queries the in-memory plugin registry where register_provider_profile() registers AntigravityProfile.

resolve_provider_full(): Used by hermes model, TUI /model, and switch_model(). It queries:

Config file (providers: block in config.yaml)

Built-in aliases (get_provider)

Saved custom providers

models.dev catalog

Because resolve_provider_full() does not query the runtime plugin registry, explicit provider switches to antigravity failed.

Adding providers.antigravity to config.yaml bridges resolve_provider_full() while leaving resolve_runtime_provider() to execute via the plugin registry.

To ensure this works out-of-the-box for all users, I updated _select_model() in antigravity_provider/hermes_plugin.py to write this section automatically whenever hermes agy select runs:

4. Supporting New Models: Gemini 3.6 Flash & Wire Mapping

Google periodically updates its model lineup (e.g. introducing Gemini 3.6 Flash). However, sending a raw wire model string of gemini-3.6-flash returns 404 NOT_FOUND.

Google's API requires an explicit reasoning effort suffix on the wire:

gemini-3.6-flash-low

gemini-3.6-flash-medium

gemini-3.6-flash-high

To support google-antigravity/gemini-3.6-flash in Hermes, I updated models.py in the plugin to translate Hermes' reasoning effort settings into Google's wire model IDs:

Now, selecting google-antigravity/gemini-3.6-flash in Hermes automatically routes requests to the correct Google wire target based on your active reasoning effort knob.

5. Usage Tracking & Session Insights

A common question when integrating custom providers is whether usage analytics are preserved.

Hermes logs all token usage into ~/.hermes/state.db under billing_provider='antigravity':

Running hermes insights aggregates these tokens alongside your paid API providers.

📌 Note on Quota vs. Usage: omp includes a custom UI meter that polls Google's internal fetchUsage API to show remaining daily allowance. Hermes logs spent tokens per session locally.

6. Verification & Working Examples

Here is how to test and verify your Antigravity integration across Hermes CLI and python environments:

1. Terminal Smoke Tests

2. Programmatic Model Switching Test

7. Developer Checklist for Custom Provider Plugins

When developing or integrating custom model provider plugins into Hermes, keep this checklist handy:

Happy coding with free frontier models in Hermes! 🚀