https://github.com/kaos599/deep-ass-research
IDE-agnostic deep-research framework for AI agents: maps a topic wide, commits to depth on purpose, adversarially verifies its findings, and synthesizes them into a backlinked, primary-sourced Obsidian vault. Works in Claude Code, Cursor, opencode, or any agent.
https://github.com/kaos599/deep-ass-research
agentic-ai ai-agents anthropic claude-code claude-code-plugin claude-code-skill cursor deep-research knowledge-graph llm mcp multi-agent obsidian opencode research-agent subagents
Last synced: 5 days ago
JSON representation
IDE-agnostic deep-research framework for AI agents: maps a topic wide, commits to depth on purpose, adversarially verifies its findings, and synthesizes them into a backlinked, primary-sourced Obsidian vault. Works in Claude Code, Cursor, opencode, or any agent.
- Host: GitHub
- URL: https://github.com/kaos599/deep-ass-research
- Owner: Kaos599
- License: mit
- Created: 2026-06-22T07:06:11.000Z (11 days ago)
- Default Branch: main
- Last Pushed: 2026-06-27T12:58:58.000Z (6 days ago)
- Last Synced: 2026-06-27T19:21:40.405Z (6 days ago)
- Topics: agentic-ai, ai-agents, anthropic, claude-code, claude-code-plugin, claude-code-skill, cursor, deep-research, knowledge-graph, llm, mcp, multi-agent, obsidian, opencode, research-agent, subagents
- Language: JavaScript
- Size: 345 KB
- Stars: 3
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# 🤿 Deep-Ass Research
**Research like a paranoid PhD, not a search bar.**
Maps a topic wide, commits to depth on purpose, tries to refute its own findings, and leaves you a backlinked, primary-sourced knowledge vault — in any AI coding agent.
Before / After •
Benchmarks •
How it works •
Install •
What you get •
Full install guide
---
Most "research" an agent does is one search, three skimmed snippets, and a confident paragraph that may or may not be true. **DAR** runs the loop a careful human would: cast wide, decide where to go deep, read the **primary** source (not the thread about it), send skeptics to **disprove** every load-bearing claim, and write it all into an Obsidian vault where every sentence traces to a source you can re-open.
It's host-agnostic: the methodology is plain markdown (`core/`), and thin adapters bind it to Claude Code, Cursor, opencode, or anything that reads `AGENTS.md`.
## 🪞 Before / After
You ask: *"How much revenue does Perplexity make and what's their strategy?"*
### 🔍 Normal agent
> "Perplexity makes around $100M and is positioned as an AI answer engine competing with Google."
One search. No source. No idea if it's current, self-reported, or made up.
### 🤿 DAR
> A vault: a claim note **S-001** — *"~$100M ARR by mid-2025"* — with the **verbatim quote**, the **primary** URL, `status: verified`, an independent corroboration, and a caveat that it's *self-reported run-rate*. Linked to `[[Perplexity AI]]`, `[[Aravind Srinivas]]` (with sourced beliefs from his actual posts), and a thesis thread — plus an honest list of what's still unproven.
**Same question. One is a vibe. The other is a defensible, navigable answer you can audit.**
## Benchmarks
## How it works
DAR has two halves that share one spine — the **charter** (what we're actually trying to answer). A **gate** stops it from going deep too early; a **drift guard** stops it from wandering once it's deep. The charter is a *compass, not a cage*: useful tangents get promoted, dead ones pruned.
```mermaid
flowchart TD
A(["📥 Research request"]) --> P["0 · Preflight
bind capability verbs to host tools"]
P --> C["1 · Decompose & Charter
success criteria SC1…SCn · scope · clarify with user"]
C --> B["2 · Breadth
parallel SEARCH-only scouts map the landscape"]
B --> G{"Gate
saturated? · rankable?
anchors hit?"}
G -->|widen one wave| B
G -->|too broad: narrow| C
G ==>|human check-in| D["3 · Depth loop
divers FETCH full primaries
→ atomic, cited claim notes"]
D --> DG{"4.5 · Drift Guard
still serving the charter?"}
DG -->|CONTINUE| D
DG -->|REFOCUS: prune / promote| D
DG -->|ESCALATE: ask the user| C
DG ==>|CONCLUDE| V["4 · Verify
3 skeptics try to refute · 2-of-3 majority
→ arbiter breaks ties"]
V --> S["5 · Synthesize
librarian links & dedupes → synthesizer writes the map"]
S --> O(["📦 Deliver
Map-of-Content · report · open questions"])
classDef phase fill:#0f172a,stroke:#334155,color:#e2e8f0;
classDef gate fill:#fde68a,stroke:#b45309,color:#1c1917;
classDef done fill:#bbf7d0,stroke:#15803d,color:#052e16;
class P,C,B,D,V,S phase;
class G,DG gate;
class A,O done;
```
The principles behind each step (why they exist) — straight out of how good researchers actually work:
- **Range before depth.** Cheap, disposable `SEARCH`-only scouts first; expensive `FETCH` only on threads that survive the gate.
- **Read the primary, not the summary.** Divers fetch the full source (incl. appendices & limitations) into `raw/`; a claim sourced only to a summary is quarantined.
- **Don't fool yourself.** Skeptics are scored to *disprove* claims; disconfirming evidence is written **into** the note; contradictions are downgraded, never deleted.
- **Synthesis is the contribution.** The deliverable is a navigable `_MOC.md` + `[[wikilink]]` graph + a report answering each criterion — not a wall of text.
## Install
> [!TIP]
> **Recommended: set up these MCP servers _first_, then install DAR.** It runs on a host's built-in web tools, but is *maximized* by this free toolset — each one powers a capability verb:
>
> | MCP server | Powers in DAR |
> |---|---|
> | **Exa** | semantic / neural `SEARCH` — papers, people, companies |
> | **Tavily** | structured web `SEARCH` + `FETCH` (date-filterable) |
> | **TinyFish** | `BROWSE` — social / authenticated / interactive pages |
> | **Context7** | `READ_DOCS` — version-accurate library & API docs |
> | **Ref** | `READ_DOCS` — specs, standards, references |
>
> Set the matching keys (`EXA_API_KEY`, `TAVILY_API_KEY`, …) in your environment. On the Claude Code **plugin** path they're auto-wired from `.mcp.json` (you just enable them once). Anything missing degrades loudly, never silently — details in [`core/PREREQUISITES.md`](core/PREREQUISITES.md).
### Claude Code — plugin (one command, auto-wires MCP)
```
/plugin marketplace add Kaos599/Deep-Ass-Research
/plugin install deep-ass-research@dar-marketplace
```
Registers the seven `dar-*` role subagents, the `/deep-ass-research:dar` command, and the prerequisite MCP servers.
### Claude Code — skill (zero assembly)
```bash
git clone https://github.com/Kaos599/Deep-Ass-Research.git
ln -s "$(pwd)/Deep-Ass-Research" ~/.claude/skills/deep-ass-research
```
Then just ask for *"deep research on X"* or run `/deep-ass-research`.
### Cursor
```bash
cp adapters/cursor/deep-ass-research.mdc .cursor/rules/
cp adapters/cursor/dar.command.md .cursor/commands/dar.md
```
### opencode
```bash
cp adapters/opencode/deep-ass-research.md ~/.config/opencode/agent/
# then merge adapters/opencode/opencode.json (MCP servers + dar-* subagents)
```
### Any other agent (Codex, Aider, Gemini CLI, custom)
Load [`adapters/generic/AGENTS.md`](adapters/generic/AGENTS.md), or paste [`adapters/generic/SYSTEM_PROMPT.md`](adapters/generic/SYSTEM_PROMPT.md) into the system prompt. No sub-agents? DAR runs the roles sequentially in-context — same method.
Full per-host details: **[INSTALL.md](INSTALL.md)**.
## What you get
A self-contained Obsidian vault at `~/research/dar/--/` (override with `DAR_VAULT_ROOT`):
| In the vault | What it is |
|---|---|
| `sources/S-NNN-*.md` | one **atomic claim** per note — verbatim quote, primary URL, `status` + `confidence`, links |
| `entities/*.md` | companies / people / concepts (people carry sourced, quoted beliefs) |
| `threads/T-*.md` | synthesis notes with a stated thesis, a reasoning chain of `[[links]]`, and counter-evidence |
| `raw/` | the cached primary sources — the provenance floor, re-openable forever |
| `_verify-log.md` | every load-bearing claim's verdict + dissent |
| `_MOC.md` + `99-report.md` | the front-door map and the report answering each success criterion |
Run by seven focused roles you can spawn in parallel:
| Phase | Roles |
|---|---|
| Breadth | `scout` (cheap, wide, search-only) |
| Depth | `deep-diver` (reads full primaries) |
| Verify | `skeptic` ×3 → `arbiter` |
| Synthesize | `librarian` (links/dedupe) → `synthesizer` (the deliverable) |
| Steering | `relevance-monitor` (the drift guard) |
Repo layout
```
deep-ass-research/
├── SKILL.md # Claude Code entry (drop-in skill)
├── core/ # PORTABLE methodology — names no host tool
│ ├── capabilities.md PREREQUISITES.md methodology.md orchestration.md
│ ├── vault-layout.md note-schemas.md provenance.md
│ ├── roles/ scout · deep-diver · skeptic · arbiter · librarian · synthesizer · relevance-monitor
│ └── modes/ academic · gtm · technical · market
├── adapters/
│ ├── generic/ AGENTS.md + SYSTEM_PROMPT.md
│ ├── claude-code/ dar-pipeline.workflow.js + plugin/ (pre-assembled, installable)
│ ├── cursor/ deep-ass-research.mdc + dar.command.md
│ └── opencode/ deep-ass-research.md + opencode.json + subagents/
└── .claude-plugin/marketplace.json # makes the repo a one-command plugin
```
**Start at [`core/orchestration.md`](core/orchestration.md)** — the runbook every adapter points to.
## Research modes
DAR picks a playbook by topic, each with its own decomposition and source strategy:
| Mode | For | Move |
|---|---|---|
| **academic** | papers / literature | read methods + appendix + limitations; build a citation graph |
| **gtm** | a company, its people & beliefs | company → people → their actual posts → coherent narrative chain |
| **technical** | "how does X work", tool choices | primary docs (Context7/Ref); A-vs-B-vs-C with a recommendation |
| **market** | competitive / sector intel | players → positioning → pricing/funding/trend signals → matrix |
---
Built to be portable, auditable, and hard to fool. Read the primary. Write everything down. Don't kid yourself.