Deep research

Deep Research — Autonomous Massive Web Research

A pipeline for conducting deep, multi-source web research. Uses OpenCode's built-in websearch and webfetch tools to search, fetch dozens of pages in parallel, and synthesize comprehensive findings.

No API keys needed. No accounts. Completely free.

How It Works

User: "Search for complete info about antimatter"

Agent:
  Step 1: websearch("antimatter physics")          → 10+ source URLs
  Step 2: wefetch(source1) + webfetch(source2) + ... (parallel, 10+ at once)
  Step 3: Read & synthesize all content
  Step 4: Present structured summary with sources

Research Workflow

Step 1: Understand the Scope

Determine:

• Topic — what to research
• Depth — overview vs deep dive vs comparison
• Sources needed — general web, academic, news, documentation
• Output format — summary, comparison table, structured report

Step 2: Search for Sources

Use websearch to find relevant pages. Always search multiple queries for comprehensive coverage:

# Primary search — cast a wide net
websearch("topic")
websearch("topic research paper")
websearch("topic latest 2025 2026")
websearch("topic comparison alternatives")
websearch("topic tutorial guide")

# Collect ALL returned URLs across searches

Pro tip: Use OPENCODE_ENABLE_EXA=1 environment variable before starting OpenCode to enable websearch tool.

Step 3: Fetch Pages in Parallel

Use webfetch to fetch all collected URLs. Fetch at least 10-20 pages in parallel for proper deep research:

# Launch all fetches in parallel — do NOT wait for each one
webfetch(url1)
webfetch(url2)
webfetch(url3)
...
webfetch(urlN)

Rules:

• Fetch ALL URLs, not just a few
• Prioritize: Wikipedia, academic sources, official docs, reputable blogs
• Skip obvious low-value pages (forums, spam)
• If some fail, note it and continue

Step 4: Analyze & Synthesize

After receiving all content:

1. Identify key themes across sources
2. Note contradictions — different sources may disagree
3. Find unique insights — each source adds value
4. Structure the answer:
   • Executive summary (2-3 sentences)
   • Detailed findings with source citations
   • Key statistics / data points
   • Different perspectives
   • Conclusion

Step 5: Present Results

Format the response with:

## 📋 Research: [Topic]

**Summary:** [2-3 sentence overview]

### Key Findings
1. **[Finding 1]** — [detail] ([source](url))
2. **[Finding 2]** — [detail] ([source](url))
...

### Detailed Analysis
[Comprehensive synthesis of all sources]

### Sources
- [Title](url) — [what this source contributed]
- [Title](url) — [what this source contributed]
...

### Further Questions
- [Aspect not fully covered]
- [Related topic worth exploring]

Research Templates

Template: Comparison Research

# User: "Compare X and Y"
websearch("X vs Y comparison")
websearch("X features pricing")
websearch("Y features pricing")
websearch("X review")
websearch("Y review")

# Fetch all results in parallel
webfetch(url1) ... webfetch(urlN)

# Present as comparison table:
# | Feature | X | Y |
# | Pricing | $ | $ |
# | Pros | ... | ... |
# | Cons | ... | ... |

Template: Deep Technical Research

# User: "How does technology X work?"
websearch("X explained")
websearch("X architecture")
websearch("X tutorial")
websearch("X whitepaper")
websearch("X vs alternatives")

Template: Market / Competitive Research

# User: "Analyze the market for X"
websearch("X market size 2025")
websearch("X competitors")
websearch("X trends")
websearch("X funding")
websearch("X review")

Tips for Maximum Results

1. Search multiple queries — different angles yield different sources
2. Fetch everything in parallel — bash & style or sequential tool calls
3. Prioritize quality sources — Wikipedia, official docs, .edu, .gov, reputable tech sites
4. Note failed fetches — try alternative URLs if available
5. Cite sources — always link back to original content
6. 10+ pages minimum — for truly deep research, aim for 20-30+ pages

How This Beats Firecrawl

Feature	Firecrawl	This Skill
API Key Required	✅ Yes	❌ No
Cost	Free tier (500K credits)	Free unlimited
Search	✅ Built-in	✅ via websearch (Exa)
Markdown output	✅	✅ via webfetch
Parallel fetching	✅	✅ (agent can batch)
AI synthesis	Separate API	✅ Built into agent
Customizable	Limited	✅ Full control
Open source	AGPL	✅ MIT

Limitations

• websearch requires Exa (enabled via OPENCODE_ENABLE_EXA=1 or OpenCode provider)
• webfetch fetches one URL at a time per tool call (but agent can parallelize)
• Rate limits depend on your OpenCode provider
• Some sites block automated fetching (JavaScript-heavy SPAs)