DARE: A Markdown-Only Research Orchestrator for Autonomous AI Agents
900+ pure-markdown skills for autonomous AI research, organized as 9 freely-composable packages over a 4-layer hierarchy (Campaign → Strategy → Tactic → SOP). Non-linear orchestration with backtracking, 6 MCP integrations. The AI is the researcher — you set the direction.
At a glance
- What is it?
- The De-Anthropocentric Research Engine packages 900+ markdown skills into ten composable research packages across a four-layer Campaign/Strategy/Tactic/SOP hierarchy. It is a prompt architecture, not a runtime, and that distinction decides who should adopt it.
- Who is it for?
- Adopt DARE if you already run Claude Code or Codex against a repository and want a structured, backtracking research loop with executable spec artifacts; skip it if you need a hosted service, a stable API surface, or reproducibility guarantees.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly HTML, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The Problem DARE Claims to Solve, and Whom It Is Built For
The README opens with an argument, not a feature list: "Science is dying because the human is in the way." The stated bottleneck is not compute or data but the human deciding what to search, what to read, which gaps matter, and which ideas deserve pursuit. DARE's answer is to remove that decision point from the middle of the loop. The human supplies direction; the engine runs literature acquisition, gap discovery, hypothesis generation, adversarial stress testing, and experiment design on its own.
The intended user is therefore narrow. You need an agent harness that can execute markdown instructions at scale (the README names Claude Code and Codex in the topic list), a willingness to let an agent proceed without per-step approval, and a research question open enough that systematic coverage beats a targeted keyword search. Someone who wants a chat interface for summarising three PDFs is not the audience. Neither is a team that needs a hosted service with an SLA. DARE is a single-clone distribution of markdown files plus npm dependencies, and it assumes you are comfortable operating that stack yourself.
The Four-Layer Command Hierarchy and Why Layering Matters
The architecture borrows a military decomposition: Campaign (45+ skills) decides what to research, Strategy (200+) decides when and why, Tactic (120+) decides how to combine operations, and SOP (500+) executes single-responsibility actions. The README is explicit that each layer calls only the layer directly below it. A Strategy never touches MCP tools; a Tactic never sets research direction.
That constraint is the most interesting design decision in the repository. It means a literature-survey Strategy can hold a search-read-reflect loop, manage ledgers and budgets, and decide when to stop, without knowing whether the underlying search runs through Semantic Scholar or Brave. Swapping the retrieval backend is a SOP-level change. The cost is indirection: a single literature pass traverses four layers of markdown before anything executes, and every layer is another place for an agent to misread an instruction. The README claims every component is "independently testable, replaceable, and composable," but testability here means the instruction files can be inspected and swapped, not that there is a test suite behind them. Nothing in the supplied material describes automated tests for the skill definitions themselves.
Non-Linear Orchestration: Campaigns Compose Freely, With Backtracking
The ten campaigns are north-star-crystallization, knowledge-acquisition, deep-insight, hypothesis-formation, creative-ideation, convergence, stress-test, experiment-execution, knowledge-structuring, and ara-from-context. The README states they are "freely composed (no fixed order)" and that each owns a complete research phase with its own completion criteria, backtrack conditions, and context protocol.
This is where DARE diverges from a pipeline tool. A linear pipeline assumes the output of stage N is good enough for stage N+1. DARE instead lets a stress-test campaign send work back to hypothesis-formation, and the README's release history reflects that emphasis: v3.2.2 is labelled the "falsification-first-stress-test family," and v3.2.0 added ara-from-context as a tenth package. The mechanism for carrying state across those jumps is the Executable Research Spec, described as a machine-readable document with checkbox progress tracking, quantified completion criteria, backtrack conditions, and session recovery, so that another agent instance can resume where the previous one stopped.
That is a concrete artefact, not a slogan, and it is the part most likely to survive scrutiny. What the README does not specify is the schema of that spec. Without a documented format, resuming across agent instances depends on the model reading the same file the same way twice.
Getting It Running: One Clone, One npm install, Several MCP Servers
The install path is deliberately short. The repository is a single-clone distribution: every skill declares its dependencies inline with no external imports, so cloning the repo gets the full skill set. The README states that the custom MCP servers semantic-scholar-mcp and wiki-vault are published as npm packages and declared as dependencies, so npm install pulls them in. It does not give an explicit command sequence beyond that, so the practical steps are clone, then npm install at the repository root.
Configuration lives in a Configuration section of the README that was truncated in the supplied material, so the exact config keys are not verifiable here. What is verifiable is the integration surface: the README lists seven MCP integrations, Semantic Scholar, Brave Search, Tavily, Keenable, AlphaXiv, Apify web scraping, and Wiki Vault. Each of those except the two Yogsoth-built servers is a third-party service, which means API keys and, for several of them, paid quotas. The README does not document rate limits, cost per research run, or what happens when a key is missing. Treat the seven integrations as an upper bound you configure down to what you can pay for, not as a default working set.
Where the Markdown-Only Design Breaks Down
The primary language of the repository is HTML, and the skills are pure markdown. There is no compiled orchestrator, no scheduler, and no runtime that enforces the layering rule. The four-layer hierarchy is a convention the agent is instructed to follow, not a constraint the system can reject. An agent that decides mid-run to call an MCP tool from a Strategy layer will not be stopped; it will simply produce a run whose trace no longer matches the architecture diagram.
A second limitation is reproducibility. The README promises systematic coverage through citation chaining and snowball sampling, but the same research question run twice will not necessarily traverse the same papers, because the search and reading steps are model-driven. For exploratory work that is acceptable. For anything you intend to publish as a reproducible method, it is not, and DARE does not appear to offer a deterministic mode.
Third, the philosophical framing is load-bearing. The README argues that human desire is mimetic, that institutions filter for conformity, and that the human's role should shrink to oracle and guardian. If you disagree with that premise, you are not just disagreeing with the introduction; the campaign design assumes the agent proceeds "without asking for permission," which is exactly the behaviour a more cautious research lead would want to gate. There is no documented approval checkpoint between hypothesis-formation and experiment-execution in the supplied material.
How DARE Differs From a Conventional AI Research Assistant
The obvious comparison is a general-purpose deep-research assistant of the kind bundled with major model providers: you type a question, it runs a bounded number of searches, and it returns a synthesised report. That tool is optimised for a single answer delivered once. Its loop terminates when the report is written.
DARE inverts the unit of work. The output is not a report but an Executable Research Spec with checkbox progress and backtrack conditions, and the campaigns can re-enter earlier phases when stress testing falsifies a hypothesis. The README's stress-test campaign includes sacred cow hunting, assumption destruction, and worst-case design, with the stated rule that "ideas must survive attack before acceptance." A single-shot deep-research tool has no equivalent stage, because it has no persistent hypothesis to attack.
The trade-off is operational weight. A deep-research assistant needs a prompt. DARE needs a cloned repository, an npm install, configured MCP credentials, and an agent harness willing to run for many iterations. If your question fits in one search session, DARE is the wrong tool and will cost you more setup than it returns.
Maintenance, Versioning and the Apache-2.0 Boundary
DARE is Apache-2.0 licensed and not archived, with the last push dated 2026-09-09 and releases roughly quarterly across 2026: v3.2.0 in June, two patches the same day, and v3.2.2 four days later. That cadence suggests active maintenance, but the patch releases landing hours apart also suggest the skill definitions change quickly enough to need immediate correction. Because the skills are markdown, an upgrade is a git pull rather than a migration, which keeps upgrade cost low. The cost that does not disappear is re-validation: after a pull, the behaviour of a campaign can change without any version number signalling which skill file moved.
On licensing, Apache-2.0 permits commercial use and modification and includes a patent grant, which matters if you build on the skill definitions. It also requires that you preserve notices and state changes. This is a description of the licence text, not legal advice; if you plan to redistribute a modified skill set, have counsel review the NOTICE requirements.
The dependency chain is the less comfortable part. npm install pulls semantic-scholar-mcp and wiki-vault from the Yogsoth organisation, so your reproducibility depends on two small packages maintained by the same author as the skills. Pin them if you care about a run being repeatable six months from now.
Editorial conclusion
Adopt DARE if you already run Claude Code or Codex against a repository and want a structured, backtracking research loop with executable spec artifacts; skip it if you need a hosted service, a stable API surface, or reproducibility guarantees. Before committing, verify that the two npm MCP servers (semantic-scholar-mcp, wiki-vault) resolve on your Node version, and read assets/DE-ANTHROPOCENTRIC.md to confirm you accept the philosophical framing, since it is not separable from the skill design.
Community notes