Comparison with Alternatives

Comparison with Alternatives

Where the kit sits in the 2026 landscape of repository documentation: hosted AI wikis, self-hosted generators, static-site pipelines, and the baseline it most often replaces — an agent told to "write a wiki for this repo" with no tooling. The comparison is by mechanism, not by marketing: who does the research, whether claims get verified, where the output lives, and what it costs. External facts are cited to their sources; kit facts are traced to code or to this run's tests.

Contents

• The four approaches
  • 1. Hosted AI wikis — DeepWiki and friends
  • 2. Static-site documentation pipelines — MkDocs, Docusaurus and kin
  • 3. The baseline — an agent with no tooling
  • 4. This kit — agent-operated, verification-first
• Matrix
• Choosing
• See also

---

The four approaches

1. Hosted AI wikis — DeepWiki and friends

DeepWiki (Cognition, launched May 2025) turns any public GitHub repository into browsable documentation with architecture diagrams, source links and a chat interface — free for public repos, reached by swapping github.com for deepwiki.com in a URL; configuration lives in a .devin/wiki.json, and private repositories are served through Devin. Open-source alternatives such as DeepWiki-Open run the same idea locally against any git host; commercial 2026 offerings (e.g. Repowise's private-repo comparison surveys the field) compete on privacy, freshness and private-repo support.

Strengths: zero effort, always available, chat is a genuinely useful interface, diagrams without work.

Limits, by mechanism: the documentation is generated from reading the repository (retrieval + generation). Nothing executes the subject, so behavioural claims carry exactly the README's reliability — the failure mode this kit exists to prevent. Output lives on the vendor's host (or your local DeepWiki-Open instance), not in your repository's own wiki; you cannot edit, review as prose, or ship it as a git artefact; and regeneration replaces rather than revisions.

2. Static-site documentation pipelines — MkDocs, Docusaurus and kin

SSGs are the established route for versioned, reviewed, self-hosted docs: Markdown-in, site-out, deployed beside the code, diffable in PRs.

Strengths: total ownership, mature themes, versioning, search, plugin ecosystems, docs-as-code review flow.

Limits, by mechanism: an SSG renders what a human (or agent) wrote — it contributes nothing to research or verification. The blank-page problem, the README-transcription failure mode and the QA burden are all still yours. Different problem class: rendering and hosting, not truth-finding. (A kit-built wiki and an SSG can coexist: the research files and verified tables port directly into SSG pages.)

3. The baseline — an agent with no tooling

The control group: paste "write a complete wiki for this repo" into a coding agent. This is precisely the scenario the kit's reference run exposed: the agent transcribes the README's syntax highlighting and nested lists claims — both false — because nothing in the loop runs the software, checks anchors, audits links, or records which statements were verified. Output quality becomes a function of the model's guesswork and the reviewer's patience.

The kit is this baseline plus the machinery that turns guesses into evidence: staged execution, declarative batteries, a fact ledger with per-value provenance, structural auditing with self-tests, ban lists built from claims deliberately not made, and a delivery pipeline that survives a credential-less sandbox.

4. This kit — agent-operated, verification-first

Mechanism: an LLM agent walks four phases; tools scan, stage, probe, scaffold, audit and package; files are the interface; every claim carries a source. Output: GitHub's native wiki (.wiki.git, branch master) as a commit + verified bundle, or any directory of Markdown you prefer.

Matrix

Dimension	DeepWiki (hosted)	DeepWiki-Open (local)	MkDocs / Docusaurus	Agent, no tooling	This kit
Who researches	vendor pipeline (reads repo)	your LLM (reads repo)	you	the model's recall	the agent, with tools that run the subject
Claim verification	none published	none by default	manual review	none	one rule: source line or reproduced execution; facts.json provenance per value
Empirical probing	no	no	no	no	20 traversal + 26 injection payloads, headers, flags, uploads, disk (0.46 s measured, T14)
False-claim defence	—	—	reviewer	—	ban.json from sources §D, auditor-enforced with negation-aware matching
Structural QA	vendor-side	—	link checkers (per plugin)	—	16 check families; anchors via GitHub's slug algorithm; bash -n/compile() on code blocks; self-test proves the checker can fail
Output location	vendor host	local web UI	your site	chat/files	GitHub native wiki via commit+bundle; plain Markdown directory
Ownership/editability	read-mostly	yours (regenerate)	yours	yours	yours — git artefacts, human-editable prose
Cost model	free (public repos); Devin for private	your LLM + compute	your time + CI	model tokens	model tokens + ~0 infra (zero dependencies)
Privacy of the subject	code processed by vendor	local	local/CI	depends on the model provider	local: only apt/git egress; probes hit 127.0.0.1
Freshness	regenerated	regenerated	per commit	per ask	per run; pinned to a commit SHA in the footer
Effort to first result	seconds	minutes	hours	minutes	a run: research ~40%, writing ~45% of the budget
Recovery from interruption	n/a	n/a	n/a	restart	STATE/PROGRESS/REWORK-STATE + snapshots; reference run resumed at zero cost

Choosing

Situation	Reasonable choice
Public repo, casual browsing, zero budget	DeepWiki — seconds to value, chat interface
Private code, no data may leave the machine, casual docs	DeepWiki-Open or an SSG with human authorship
Long-lived product docs under PR review, versioned per release	An SSG (MkDocs/Docusaurus) — optionally seeded by a kit run's verified tables
A wiki that must be trustworthy about behaviour — security properties, exact status codes, config effects, doc-vs-code discrepancies	this kit: it is the only option in this list whose mechanism executes the subject and enforces provenance
Existing GitHub wiki you want to complete/audit	this kit — the auditor works on any directory of wiki Markdown, and delivery targets .wiki.git directly

Honest statement of the kit's costs: it needs a capable tool-using agent (the tools do not write prose); a full run is hours of agent work, not seconds; the probe stack assumes an environment where the subject can be executed; and the 20-page reference composition is a budget, not a button. What it buys is the property none of the alternatives provide by mechanism: a wiki where every behavioural claim was either read at a line number or reproduced by a named test — and where the claims that failed verification are recorded, banned and still documented as discrepancies.

Note

Vendor capabilities change quickly. The DeepWiki facts above were current as of its launch materials (May 2025) and the 2026 private-repo surveys cited; re-check before making a procurement decision. Kit-side facts are pinned to commit 7e17a11 and this run's tests.

---

See also

• Features and Verification Model — the mechanism in detail
• Home — what the kit is, in two paragraphs
• FAQ — "can I use this on private repos / non-PHP subjects / an existing wiki?"
• Publishing and Delivery — where the output lands and how
• Glossary — canonical terms used across this wiki