De-Anthropocentric Research Engine: 900+ Markdown Skills That Run Research as a Command Hierarchy

What DARE Actually Solves, and Who It Is Written For

The README states the problem in blunt terms: existing AI research assistants still require a human to decide what to search, which gaps matter, and which ideas deserve pursuit. DARE's answer is to move the human out of the center of that loop. The human supplies an initial direction; the system is supposed to handle literature acquisition, gap discovery, hypothesis formation, stress testing and experiment design on its own.

The audience is narrow and identifiable. This is not a tool for a biologist who wants a summary of five papers. It is for people already running an agentic coding environment, because the skills are markdown files meant to be read and executed by a model, and the repository carries a dsh-plugin directory alongside a cli directory. If you do not already have an agent that can consume skill files, DARE gives you a well-organized tree of instructions and nothing that executes them for you.

The README's framing is ideological as much as technical. It cites a 90% decline in scientific disruptiveness since 1945 (Park et al., 2023) and argues that researcher headcount grew while output quality fell. That argument is the justification for the architecture, not a feature of it. You can use the skill tree without accepting the philosophy, but the philosophy explains why the layers are ordered the way they are.

Campaign, Strategy, Tactic, SOP: The Four-Layer Command Structure

The architecture is a strict hierarchy, and the README is explicit that each layer calls only the layer directly below it. A Strategy never touches MCP tools. A Tactic never decides research direction. The README gives a military analogy for the decomposition:

Campaign (45+)  →  "Take that hill"         →  WHAT to research (full research stage)
Strategy (200+) →  "Flank from the east"    →  WHEN and WHY (iteration loops, stopping conditions)
Tactic (120+)   →  "Squad A cover, B move"  →  HOW to combine (orchestrates multiple SOPs)
SOP (500+)      →  "Fire, reload, advance"  →  HOW to execute (single-responsibility operations)

The counts matter for judging scope. Roughly 500 SOPs sit at the bottom doing single-responsibility work, 120 tactics combine them, 200 strategies hold state such as ledgers and budgets and decide when to stop, and 45-plus campaigns own complete research phases. The README names ten campaigns: north-star-crystallization, knowledge-acquisition, deep-insight, hypothesis-formation, creative-ideation, convergence, stress-test, experiment-execution, knowledge-structuring, and ara-from-context.

Two design decisions stand out. First, campaigns are freely composed with no fixed order, which is what the project means by non-linear orchestration with backtracking. Second, the skills are pure markdown with dependencies declared inline and no external imports, which is why one clone is meant to be sufficient. That self-containment is the real engineering claim here: a skill file that needs no import resolution can be read by any agent that can read text.

Installing DARE and Running a First Campaign

The README describes this repository as the single-clone distribution, so the first step is a clone rather than a package install. The package.json sets engines to Node 22.0.0 or higher, which is the constraint to check before anything else.

git clone https://github.com/yogsoth-ai/de-anthropocentric-research-engine.git
cd de-anthropocentric-research-engine
npm install

The README states that the repository declares its custom MCP servers as dependencies so that npm install pulls everything needed. The package.json confirms five runtime dependencies: @apify/actors-mcp-server, @brave/brave-search-mcp-server, @yogsoth-ai/semantic-scholar-mcp, @yogsoth-ai/wiki-vault, and tavily-mcp. Note that the root package is marked private, so npm install installs dependencies rather than publishing or globally linking DARE itself.

Configuration follows the MCP convention. The repository ships an mcp.example.json at the top level, which is the file to copy and fill in with your own credentials for the search and scraping servers. The README does not document the exact key names inside mcp.example.json, so read that file and edit it in place rather than guessing.

For a first real run, the README's entry point is direction crystallization: you give the system a cold start from zero, a warm start from a vague interest, or a hot start from a specific question, and the north-star-crystallization campaign produces a structured North Star. The repository also ships an assets/DE-ANTHROPOCENTRIC.md file holding the longer philosophical argument, and a docs/ directory that is the place to look for per-campaign detail the README does not carry.

Where the Autonomy Claim Breaks Down

The README says DARE works without asking for permission. The honest reading is that autonomy here means the skill files do not include approval gates, not that the system is free of human input. You still supply the direction, you still supply API keys for Semantic Scholar, Brave, Tavily, Keenable, AlphaXiv, Apify and Wiki Vault, and you still decide when a campaign's completion criteria have been met. The README itself assigns the human two roles, oracle and guardian, and describes the guardian as maintaining ethical floors and sanity checks. That is a human in the loop by another name.

The harder limitation is verification. The README claims 900+ skills and gives layer counts of roughly 500 SOPs, 120 tactics, 200 strategies and 45 campaigns. Those are counts of files, not evidence that any given campaign produces correct research. The falsification-first stress-test family arrived in v3.2.2, and the naming suggests the project is aware that generated hypotheses need adversarial treatment, but the README does not describe how a user audits whether a stress test actually attacked the right assumptions.

There is also a structural mismatch worth naming. The repository's primary language is listed as HTML, while the substance is markdown skills and JavaScript tooling. Anyone expecting a conventional HTML application from that label will be looking in the wrong place. The cli/ and dsh-plugin/ directories are where executable behavior lives, and the README does not document their command surface.

How DARE Differs from a Literature Review Assistant

The natural comparison is a literature review tool that takes a query, retrieves papers, and returns a summary. The difference is architectural rather than incremental. A review assistant treats retrieval as the whole task. DARE treats retrieval as one campaign among ten, sitting between north-star-crystallization above it and deep-insight below it, with its own strategies for managing the search-read-reflect loop and its own saturation detection.

That changes what the output is. The README describes Executable Research Specs as machine-readable documents with checkbox progress tracking, quantified completion criteria, backtrack conditions and session recovery, and states that another Claude Code instance can pick up where you left off. A review assistant produces a report; DARE aims to produce a resumable process artifact. Whether that distinction pays off depends on whether your work is a one-off question or a multi-session investigation.

The cost of the difference is setup. A hosted review tool works from a browser. DARE requires a clone, Node 22, npm install, an MCP configuration file with credentials for up to seven servers, and an agent environment that can read markdown skills. If your literature question can be answered in one sitting, that setup is not worth it.

Maintenance, Licensing and the Upgrade Path

The repository is not archived, and the last push was on 2026-09-09, so the tree is recent. The release cadence visible in the notes is uneven rather than steady: v3.2.0 on 2026-06-17 introduced ara-from-context as a tenth package, v3.2.1 and v3.2.2 both landed on 2026-06-21, with v3.2.2 carrying the falsification-first stress-test family. Between late June and early September the repository saw pushes without a tagged release, which suggests work continued on the tree before the next version was cut.

Upgrade cost depends on how you have customized the skills. Because the skills are pure markdown with inline dependency declarations and no external imports, editing a skill file is easy and merging upstream changes into that file is correspondingly awkward. There is no documented migration mechanism for local skill modifications, and the README does not document rollback. If you fork skills, track your edits yourself.

The five MCP dependencies are version-ranged in package.json, which means npm install can move them forward within their ranges. If a Semantic Scholar or Tavily server changes behavior, the failure will surface inside a campaign rather than at install time.

Licensing is Apache-2.0 per both the repository metadata and the package.json license field. Apache-2.0 includes an express patent grant and requires that you retain notices and state significant changes when redistributing. That last requirement is the one to watch if you fork skills and ship them internally. This is a description of the license text, not legal advice; check with counsel if you plan to redistribute a modified tree commercially.

Editorial conclusion

Adopt DARE if you already run Claude Code or Codex and want a structured, resumable research process with literature search wired in through MCP. Do not adopt it if you want a turnkey application: there is no compiled binary, no hosted service, and the repository's own package.json marks the root package as private, so the entry point is the skill tree and the CLI directory rather than an npm package. Before committing, verify two things: that your Node version is 22.0.0 or higher, since engines enforces it, and which of the five declared MCP dependencies you actually have credentials for, because npm install will pull servers you cannot call without keys. The repository was last pushed on 2026-09-09, so the tree is current, but the README's own roadmap section is where you should check whether the campaign you need is finished.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. README
4. Releases
5. yogsoth-ai/de-anthropocentric-research-engine on GitHub