Tota Agent

Once you're Tota, you'll never be OpenClaw.

Tota Agent is a Brazilian-fast fork of Hermes Agent, tuned for low-latency JSON, faster async I/O, typed tool-call parsing, and Rust-ready hot paths. It keeps the Hermes Agent operating model while giving this fork its own brand, benchmark story, and public launch page.

The visual identity is inspired by Tota MC's public Brazil-to-US streaming rise: creator energy, Rocinha-to-global momentum, improvised live culture, and cross-language charisma. Public references include the Streamer University coverage by Times of India and the Portuguese profile syndicated by Rede NXT. The core geometric logo does not use a portrait or imply official endorsement; the benchmark battle cards also include the supplied circular Tota mark for campaign use.

Launch Assets

• Standalone HTML site
• Tota vs OpenClaw launch banner PNG
• Tota vs OpenClaw launch banner SVG
• Benchmark battle cards
• Hermes 0.14.0 side-by-side report
• Daily Hermes sync routine
• Updated benchmark PDF - May 18, 2026 edition with the Hermes 0.14.0 refresh, brand, site, visuals, and current .venv validation.
• SVG logo
• PNG logo
• Open graph image
• GPT-image-2 emblem source

Why Tota Agent

Need	Tota Agent answer
Keep Hermes compatibility	Forks Hermes Agent instead of replacing its architecture.
Reduce message hot-path cost	Uses the orjson/msgspec/Rust-ready direction measured in the benchmark.
Improve async responsiveness	Uses the uvloop direction for Python I/O scheduling where supported.
Tell a sharper product story	Adds Tota Agent branding, launch site, and benchmark visuals.
Compare against alternatives	Includes measured comparisons with Hermes Original and OpenClaw.

Install

From GitHub

git clone https://github.com/wesleysimplicio/tota-agent.git
cd tota-agent

uv venv .venv --python 3.11
source .venv/bin/activate
uv pip install -e ".[all,dev]"

./hermes

Windows users can use the native PowerShell installer at scripts/install.ps1.

From This Checkout

cd /Users/wesleysimplicio/Projetos/contribuicoes/hermes/tota-agent
source .venv/bin/activate 2>/dev/null || source venv/bin/activate
uv pip install -e ".[all,dev]"
./hermes

Performance Extras

The benchmarked Tota Agent direction is built around fast Python plus native-extension-ready hot paths:

uv pip install -e ".[fast]"

Build the Rust extension and verify the native fast path:

PATH="$HOME/.cargo/bin:$PATH" bash scripts/install-rust.sh
python -c "from agent._hermes_fast import HAVE_RUST; print('Rust:', HAVE_RUST)"

The fast extra stays optional so the base install remains small. When present, Tota Agent uses orjson, msgspec, uvloop, and the Rust extension with Python fallbacks for locked-down or source-only environments.

Daily Hermes Sync

Tota Agent can run a daily sync routine that updates the local environment, runs hermes update, merges the latest NousResearch/hermes-agent core, and keeps Tota's speed customizations under validation before pushing a dated branch:

python3 scripts/install_tota_hermes_daily_update_launchd.py --hour 6 --minute 30

See docs/tota-hermes-daily-update.md.

Post-Benchmark Performance Patch

Version 0.14.2 adds the Hermes 0.14.0 side-by-side benchmark refresh, the daily Hermes sync routine, and the report generation dependency needed to regenerate tota_agent_benchmark_report.pdf.

Version 0.13.3 keeps the local validation path reliable: the canonical scripts/run_tests.sh runner now works when called without arguments, and the ACP registry manifest is pinned to the same package version as pyproject.toml.

Version 0.13.2 keeps the benchmark follow-up patch and switches the Tota fork's default home from ~/.hermes to ~/.tota for new installs. TOTA_HOME is the fork-native override, while HERMES_HOME remains supported for existing hermes2 deployments such as ~/.hermes2.

Version 0.13.1 applied the benchmark follow-up plan:

• Bytes-native JSON via agent._fastjson.dumps_bytes() for short payload hot paths.
• Direct Rust serde_json::Value to Python object conversion for tool-call deltas.
• Batched token helpers: estimate_tokens_many() and estimate_messages_tokens().
• Rust bytes variants for message-token estimation/truncation.
• Automatic uvloop policy installation in CLI and gateway entrypoints when available.
• Bounded fast extra dependencies to keep supply-chain risk controlled.

Details: docs/tota-benchmark-win-plan.md.

Benchmark Headline

Metric	Hermes Original	Tota Agent	OpenClaw	Winner
Total score	30 / 50	44 / 50	36 / 50	Tota Agent
JSON dumps, large payload	18.40 us	3.20 us	5.80 us	Tota Agent
JSON loads, large payload	12.80 us	2.80 us	5.20 us	Tota Agent
Medium message pipeline	7.50 us	2.20 us	3.46 us	Tota Agent
Medium message throughput	133k msg/s	454k msg/s	289k msg/s	Tota Agent
Tool-call typed parse	Error / N/A	0.45 us	N/A	Tota Agent
Async 1,000 tasks	2.50 ms	1.40 ms	0.08 ms	OpenClaw
Cold start	~52 ms	~50 ms	~280 ms	Tota Agent
RSS memory	~30 MB	~30 MB	~97 MB	Python variants

The repo also ships a dedicated side-by-side harness for upstream stock Hermes 0.14.0: scripts/benchmark_tota_vs_hermes_0140.py. The latest measured status lives in docs/tota-benchmark-hermes-0.14.0.md and was folded into the refreshed PDF.

Benchmark source: tota_agent_benchmark_report.pdf, updated May 18, 2026 with the Tota Agent launch package, Hermes 0.14.0 side-by-side data, and current Apple Silicon .venv validation.

Benchmark Battle Cards

These shareable comparison cards turn the report's headline battles into a Tota Agent vs Hermes Agent vs OpenClaw visual campaign. They are generated by scripts/generate_tota_battle_cards.py from the benchmark values above.

Benchmark Visuals

Full Comparison Report

System Overview

Attribute	Hermes Original	Tota Agent	OpenClaw
Language	Python 3.14	Python 3.11.14	TypeScript / Node.js 22
JSON engine	stdlib json	orjson	V8 built-in JSON
Event loop	asyncio	uvloop	libuv
Struct decode	None	msgspec	None
Native extension	None	Rust / PyO3 ready	None
Channels measured	WhatsApp, HTTP	WhatsApp, HTTP	WhatsApp, Telegram, Discord, HTTP
Channels in current checkout	WhatsApp, HTTP	Telegram, Discord, Slack, Matrix, Signal, email, SMS, API server, and more	WhatsApp, Telegram, Discord, HTTP
Category	AI Agent	Optimized Python AI Agent	Multi-channel AI Gateway

Architecture

Component	Hermes Original	Tota Agent	OpenClaw
Runtime	CPython 3.14	CPython 3.11.14	Node.js 22 / V8
HTTP client	httpx / aiohttp	httpx + uvloop	axios / undici
JSON	stdlib json	orjson 3.x	V8 JSON
Streaming	SSE asyncio	SSE uvloop optimized	SSE libuv
Tool calls	json.loads	Rust ext + orjson + msgspec	JSON.parse
Tokens	naive len // 4	Rust-ready estimate_tokens()	JS split
Packaging	pip / venv	pip / venv + Rust .so	npm / node_modules

JSON Serialization

Lower latency is better.

Payload	Hermes dumps	Tota dumps	OpenClaw dumps	Tota vs Hermes
Short, ~50 B	1.29 us	0.21 us	0.17 us	6.1x faster
Medium, ~600 B	3.38 us	0.80 us	1.00 us	4.2x faster
Large, ~50 KB	18.40 us	3.20 us	5.80 us	5.8x faster

Payload	Hermes loads	Tota loads	OpenClaw loads	Tota vs Hermes
Short, ~50 B	0.62 us	0.30 us	0.33 us	2.1x faster
Medium, ~600 B	2.90 us	1.30 us	2.29 us	2.2x faster
Large, ~50 KB	12.80 us	2.80 us	5.20 us	4.6x faster

Memory

Metric	Hermes Original	Tota Agent	OpenClaw
json.dumps medium heap / 1k calls	~420 KB	~180 KB	~160 KB
json.loads medium heap / 1k calls	~380 KB	~140 KB	~200 KB
msgspec encode medium heap / 1k calls	N/A	~95 KB	N/A
Process RSS	~30 MB	~30 MB	~97 MB
Disk footprint	~10 MB	~15 MB	~200 MB

Message Pipeline

Pipeline metric	Hermes Original	Tota Agent	OpenClaw	Tota vs Hermes
Short message latency	2.10 us	0.55 us	0.55 us	3.8x faster
Medium message latency	7.50 us	2.20 us	3.46 us	3.4x faster
Short message throughput	476k msg/s	1.82M msg/s	1.82M msg/s	3.8x
Medium message throughput	133k msg/s	454k msg/s	289k msg/s	3.4x

Tool-Call Parsing

Method	Hermes Original	Tota Agent	OpenClaw
JSON parse path	ERROR	1.30 us	0.54 us
orjson.loads	N/A	1.00 us	N/A
msgspec ToolCall struct	N/A	0.45 us	N/A
Rust parse_tool_call_delta	N/A	~0.40 us	N/A
Throughput	N/A	~2.5M/s	~1.85M/s

Tokens, Async, Startup

Metric	Hermes Original	Tota Agent	OpenClaw	Winner
Fast token estimate	0.12 us	0.10 us	0.04 us	OpenClaw
Token throughput	8.3M texts/s	10M texts/s	25M texts/s	OpenClaw
1,000 async tasks	2.50 ms	1.40 ms	0.08 ms	OpenClaw
Async batches/s	400/s	714/s	12,500/s	OpenClaw
Cold start total	~52 ms	~50 ms	~280 ms	Tota Agent

Category Score

Category	Hermes Original	Tota Agent	OpenClaw
JSON performance	2 / 5	5 / 5	4 / 5
Memory	5 / 5	5 / 5	2 / 5
Message throughput	2 / 5	5 / 5	4 / 5
Tool-call parsing	1 / 5	5 / 5	4 / 5
Token counting	3 / 5	3 / 5	4 / 5
Concurrency / async	3 / 5	4 / 5	5 / 5
Startup / cold start	4 / 5	5 / 5	2 / 5
Integrations	3 / 5	3 / 5	5 / 5
Library ecosystem	2 / 5	5 / 5	4 / 5
Disk footprint	5 / 5	4 / 5	2 / 5
Total	30 / 50	44 / 50	36 / 50

Usage Recommendations

Scenario	Recommended	Reason
WhatsApp / HTTP AI agent	Tota Agent	4-6x faster JSON path with Hermes-compatible Python ergonomics.
Serverless / Lambda / Cloud Run	Tota Agent	~50 ms cold start vs ~280 ms for OpenClaw.
Low memory footprint	Tota Agent	~30 MB RSS vs ~97 MB for OpenClaw.
Existing Python production stack	Tota Agent	Drop-in optimized fork direction.
1,000+ concurrent connections	OpenClaw	Native libuv scheduler wins pure scheduling benchmarks.
Multi-channel out of the box	Tota Agent	The current checkout includes more gateway adapters than the benchmarked Tota subset.
Hermes upstream contribution baseline	Hermes Agent	Canonical upstream project and community.

Development

source .venv/bin/activate 2>/dev/null || source venv/bin/activate
python -m pytest
python -m ruff check .
taskflow run .

For this repository, taskflow inspect . detects the Python and Node surfaces and taskflow run . produces the local validation checklist.

Upstream

Tota Agent is a fork of NousResearch/hermes-agent. The upstream project provides the core Hermes agent architecture, CLI, gateway, tools, skills, sessions, and multi-platform agent runtime. This fork adds a Tota Agent brand layer, benchmark campaign, performance-oriented packaging story, and launch site.