Summary

A self-contained, zero paid-API contribution that turns Hermes into a citation-disciplined research agent on the latest open-source stack. Two pieces, one branch, no modifications to existing code:

1. tools/local_web_tools.py — free-tier counterpart to web_tools.py

Hermes' existing web_search / web_extract rely on Firecrawl, Parallel, Tavily, Exa, Gemini — all paid. This file mirrors their JSON contracts using free local-first backends:

• Search chain (auto-fallback): SearXNG → Brave free tier → Tavily free tier → ddgr → ddgs/duckduckgo_search
• Extract: lynx -dump with boilerplate stripping (nav menus, button labels, iframe markers, captcha blocks, cookie notices)
• Summarization (optional): any local OpenAI-compat endpoint at $LLM_BASE_URL

2. skills/research/deep-research/ — methodology skill

Pure markdown skill teaching a 5-phase research pipeline with strict citation discipline:

1. Decompose topic → 4-6 sub-questions
2. Fan-out search across sub-questions
3. Fetch promising URLs selectively
4. Cross-verify claims; assign confidence stars (★★★ / ★★ / ★ / ⚠ / ?)
5. Synthesize structured report with mandatory Open-Questions section

Backend-agnostic — works equally with paid web_search/web_extract OR the new free local_web_search/local_web_extract.

Anti-fabrication built in: every quantitative claim needs [n] citations, Open-Questions section is mandatory, post-process verifier (shell helper in SKILL.md) flags un-cited numbers and dangling references.

Multi-backend LLM support ($LLM_BASE_URL)

Configurable, no code changes:

LLM_BASE_URL=http://localhost:11434     # Ollama (default)
LLM_BASE_URL=http://localhost:8088      # llama.cpp's llama-server
LLM_BASE_URL=http://localhost:8000      # vLLM
LLM_BASE_URL=http://localhost:1234      # LM Studio

First-class Qwen3.5 / Qwen3.6 support

The recommended local stack uses the latest Qwen open-source releases (Apache 2.0):

Critical operational note baked into the SKILL.md and tool: Qwen3.5/3.6 do NOT honor /think /no_think directives the way Qwen3 did. Per the official model card, the only reliable way to disable thinking is chat_template_kwargs.enable_thinking=false at the request level. local_web_tools.py auto-detects Qwen3.5/3.6 model tags via _is_qwen35_or_36() and applies this flag automatically.

Includes scripts/start-llama-server.sh — turnkey launcher that auto-detects Qwen3.5/3.6 from the GGUF filename and applies sane defaults (--jinja, port 8088, ctx 16384, configurable N_GPU_LAYERS).

Quickstart (5 commands, zero paid keys)

# 1. Get llama.cpp prebuilt
curl -fsSL "https://github.com/ggerganov/llama.cpp/releases/download/b9010/llama-b9010-bin-ubuntu-x64.tar.gz" | tar xz

# 2. Pull a Qwen3.5 GGUF
curl -fLO "https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/main/Qwen3.5-4B-Q4_K_M.gguf"

# 3. Boot llama-server
~/.hermes/skills/research/deep-research/scripts/start-llama-server.sh ./Qwen3.5-4B-Q4_K_M.gguf

# 4. (Optional) self-host SearXNG
docker run -d -p 8888:8080 searxng/searxng

# 5. Configure
export LLM_BASE_URL=http://127.0.0.1:8088
export SEARXNG_URL=http://127.0.0.1:8888

Validation

End-to-end integration test against Qwen3.5-4B-Q4_K_M on llama-server b9010 (May 2 2026 release):

Step 6 of the agent loop hit the integration-test 180s timeout under CPU prompt-eval on accumulated context — production deployment with GPU acceleration or a more generous timeout completes the full synthesis pass. The plumbing of all four pieces is fully validated.

Why this isn't a duplication

web_tools.py is excellent for users with paid keys. This PR makes Hermes Agent's web research fully zero-cost for users who self-host. Same JSON contracts mean any skill calling web_search/web_extract works identically with the local variants — no skill-side changes.

Files

• tools/local_web_tools.py (552 lines)
• skills/research/deep-research/SKILL.md (378 lines)
• skills/research/deep-research/scripts/start-llama-server.sh (106 lines)

Total: 1,036 lines additive, zero modifications to existing files. MIT license, no new dependencies (uses existing requests, optional lynx / ddgr / ddgs already common in research environments).

Commits

• 6ff296a feat(tools): add local_web_tools — free-tier counterpart to web_tools
• 7ff5fb8 feat(tools/local_web_tools): first-class Qwen3.5/3.6 + multi-backend support
• fac1501 feat(skills/research): add deep-research methodology skill
• bcdb4f0 feat(skills/research/deep-research): full Qwen3.5/3.6 + llama.cpp first-class support

Happy to split into separate tool + skill PRs if maintainers prefer reviewing them independently.