Skip to content
GRPNR.

R4 — AI Codebase Documentation Tools & Docs-for-Agents Conventions (2025-2026)

Research date: 2026-07-18. Compiled for plans/003-atlas-intelligence-layer. Question under test: for a 1.05M-LOC EOL Ruby/Java legacy estate + a greenfield TypeScript monorepo, consumed mainly by 64 agents + 2 humans, is there a documentation TOOL worth adopting, or do agents just read markdown-in-git, making doc tooling freight?

1. DeepWiki (Cognition)

  • Launched April 25, 2025 by Cognition (maker of Devin). Free tier auto-generates a wiki for any public GitHub repo by swapping github.com → deepwiki.com. By 2026 it had indexed 50,000+ of the most-starred public repos, “billions of lines of code.” Source: cognition.com/blog/deepwiki; codersera.com 2026 guide.
  • Pricing 2025→2026: retired original Core/Team plans, now Free/Pro/Max/Teams/ Enterprise. Private-repo indexing + stronger models = usage-based billing. Enterprise tier: VPC deployment, SSO, admin controls, org-wide repo indexing, “custom Devins.” Devin Core ACU pricing ~$2.25/ACU, $20 min. Source: codersera.com/blog/deepwiki-complete-developer-guide-2026.
  • Accuracy/staleness — negative evidence, primary sources:
    • HN thread (news.ycombinator.com/item?id=45002092, and =45011615): an LLVM contributor said results “ranged from incomplete to just plain incorrect,” omitting some of the most important compiler passes. An OSS maintainer said DeepWiki was “hallucinating pretty convincingly” — naming a struct/function for something it no longer does. A LibreOffice contributor reported DeepWiki claimed LibreOffice uses the Buck build system (false).
    • Bo Lopker, “DeepWiki and the loss of control,” blopker.com/writing/12-deepwiki/, published 2025-11-11: DeepWiki indexed his project (Codebook) without consent and invented a VS Code extension, presenting unpublished/nonfunctional code as the primary, working installation method. His core complaint isn’t just accuracy — it’s that DeepWiki auto-publishes to a URL other search engines/bots crawl, so hallucinations propagate as fact with zero human review gate. He recommends default opt-in, not automatic indexing.
    • Even DeepWiki’s own docs/third-party guides (codersera 2026, workroom-productions.com) now advise: “treat as a fast first read, not ground truth,” cross-reference security-critical modules against source, and note the tool only reads code/config/imports/repo structure — not issues, PRs, ADRs, or meeting notes (i.e., no institutional “why,” only inferred “what”).
    • Academic benchmark corroborates the accuracy ceiling: CodeWiki (arXiv 2510.24428, ACL 2026) benchmarks documentation-generation systems on CodeWikiBench and reports DeepWiki’s own closed-source baseline scoring only 64.06% quality — i.e., even Cognition’s own tool, measured independently, tops out well under 2/3 correct-and-complete by the benchmark’s rubric. CodeWiki (an academic multi-agent hierarchical-decomposition system, not a shipped product) beat it at 68.79%, with a 29.53% relative improvement on Trino (1.5M LOC, comparable scale to our legacy estate) — this is a research paper, not a product you can buy.
  • Verdict for our case: DeepWiki is built for “understand an unfamiliar public OSS repo in 5 minutes,” not for a private EOL legacy estate needing precise, auditable ground truth. Confirmed hallucination pattern (invents plausible but false integrations/build systems) is disqualifying for a legacy system where wrong answers about business logic have real cost. Not a fit; not worth the enterprise spend for our estate.

2. Swimm

  • Founded 2019 Tel Aviv, $27.6M Series A (2022). Repositioned in 2025 from a pure docs tool to an “Application Understanding / code-modernization platform”: deterministic static analysis + generative AI + senior-engineer services, explicitly targeting mainframe/legacy migrations (COBOL, CICS, PL/I) — this is close to our EOL Ruby/Java problem shape. Source: swimm.io solutions/mainframe-modernization; research.com/thectoclub 2026 reviews.
  • Core original pitch: “code-coupled documentation” with patented Auto-sync that validates/updates docs in CI, flags (doesn’t silently accept) drift. Reported 45% onboarding-time reduction in vendor-cited case studies (treat as vendor number, not independently verified).
  • Swimm’s own competitive marketing is a useful data point against pure-LLM approaches: blog post titled “A blog article is not a product. No, Claude Code can’t understand COBOL applications” (swimm.io/blog/can-claude-understand-cobol) — their pitch is that deterministic AST-level static analysis (tracing COBOL REDEFINEs, GOTO/PERFORM chains, copybook expansion) is required to avoid the hallucination failure mode documented above for DeepWiki-style pure-LLM summarization at scale.
  • Pricing: opaque, LOC-based, custom quotes; Enterprise Starter ~$28/seat/mo advertised as a floor, real deals are custom. Enterprise customer list (2025): BetterHelp, PPG, Akamai, Optum, Merck, Recursion, Fifth Third Bank — mid/large enterprises, mainframe-adjacent.
  • Weakness for our case: Swimm’s docs-first product is now secondary to selling migration services (their own reviewers note this). Good evidentiary support for “deterministic analysis beats pure LLM” but the product is sized and priced for enterprise migration engagements, not a 2-human team wanting living docs.

3. Mintlify

  • Repositioned 2025-2026 as “the knowledge platform built for agents,” not just a docs-hosting SaaS. Source: mintlify.com/ (homepage), mintlify.com/docs/ai-native.
  • Concrete AI-agent features (2026): every Mintlify-hosted page serves clean Markdown via content negotiation and at <page>.md; auto-generates llms.txt, llms-full.txt, and skill.md at site root; auto-hosts an MCP server per docs site so Cursor/Claude Code/Windsurf can query docs live during a coding task; reads a repo’s AGENTS.md to customize its own drafting-agent’s style/structure conventions; the “Mintlify Agent” watches the codebase and proposes doc updates on shipped diffs (a bot-drafts, human-approves loop — not silent autosync).
  • Reported at ~$300/mo entry per one review (ferndesk.com) — this is a customer-facing docs site product (hosting, search, branded portal), which is not what our 2-human+64-agent internal use case needs; we don’t need a public-facing portal, we need agents’ working context to be accurate.
  • Verdict: Mintlify’s specific insight worth stealing for free — auto-serve .md variants, generate llms.txt/skill.md, read AGENTS.md — are all patterns replicable with zero tooling spend in a git-native docs setup; the paid product’s value-add (hosted site, search, analytics on agent traffic) is aimed at public developer-facing docs, not our internal estate.

4. Komment.ai

  • Pitch: auto-generates in-code comments + “wikis” for legacy codebases across 90+ languages using a proprietary “Adaptive Context Cruising” context-mapping approach; targets exactly our stated problem (“streamlines maintaining legacy environments”). Source: komment.ai, ssojet.com AI doc generators roundup.
  • Free tier (“free forever,” unlimited public repos, no card) + Build/Grow/Scale paid tiers with 30-day trial; VPC self-host option (AWS/Azure/GCP) for enterprise. 2025: planned GitLab/Bitbucket adapters; migrated infra AWS→GCP in ~2 weeks with 1 engineer (cited as evidence of a small, agile team — cuts both ways: small team = thin support bench for enterprise incidents).
  • Could not find independent negative-evidence sources (G2/Slashdot/Capterra searches returned no substantive Komment-specific reviews, positive or negative — the tool appears to have very thin third-party review coverage as of 2026, which is itself a signal: low real-world adoption evidence despite marketing claims). Treat vendor claims about wiki quality/“depth” as unverified.
  • Verdict: Same generic-LLM-summarization architecture as DeepWiki (auto-generate a wiki from static analysis + LLM) with none of DeepWiki’s visible scale or independent scrutiny. No reason to prefer it over DeepWiki’s approach for our legacy estate, and the underlying failure mode (plausible hallucination on legacy business logic) is architecturally the same risk.

5. Driver AI

  • YC-backed, Austin. Emerged from stealth Oct 2024 (TechCrunch), $8M seed led by GV. Positioning is closest of the field to our exact problem: “compiler- inspired” static analysis pre-computes an exhaustive model of the codebase, then serves it to AI coding agents via MCP (not just a browsable wiki — a live queryable context source for agents). Vendor claims: 80% reduction in documentation effort, “understand and document legacy code in days instead of months,” 2x faster onboarding. Source: techcrunch.com/2024/10/08, driver.ai.
  • Caution on secondary data: aggregator sites (Tracxn/PitchBook/getlatka) showed wildly inconsistent scale figures for this same small 2024 seed startup — one aggregator claimed 567 employees and $48.8M ARR, which is implausible for an ~18-month-old, single-seed-round company and reads as algorithmically-generated aggregator noise, not a fact. Treat only the TechCrunch-sourced $8M/GV seed detail as reliable; discount the rest.
  • Verdict: Architecturally the most relevant “buy” option for legacy Ruby/Java (compiler-derived facts, not just LLM prose, served live to agents over MCP) — but it is an early-stage seed startup with no independent accuracy audits found (unlike DeepWiki, which has public HN scrutiny). Worth a pilot only if procurement risk on an 18-month-old vendor is acceptable; not a slam-dunk given no third-party verification exists yet.

6. mutable.ai — STATUS: DEFUNCT (acquired/dissolved)

  • Acquired by Alphabet/Google on 2024-12-11 (per Crunchbase acquisition record + HN discussion news.ycombinator.com/item?id=43740385). Not a pivot — by end of 2024 the website was gone, social channels dark, CEO quietly joined Google. HN commentary: the acquisition was widely read as an acquihire for an idea (“essentially doing this… acquired by Google last year”) rather than a product win. Pre-acquisition: seed-only funding, <10 employees, no Series A, losing to GitHub Copilot on adoption. Source: news.ycombinator.com/ item?id=42542512, dang.ai tool profile.
  • Verdict: Confirmed dead as an independent product/vendor. Do not evaluate further; cannot be procured.

7. CodeSee — STATUS: DEFUNCT (acquired, standalone product sunset)

  • Founded 2019, cloud codebase-visualization tool. Acquired by GitKraken on 2024-05-14 (Crunchbase acquisition record, finsmes.com). The standalone CodeSee product was sunset; its capabilities were folded into GitKraken’s “DevEx” platform. Source: koalr.com/blog/codesee-alternatives (migration guide written for displaced CodeSee customers — itself evidence the standalone tool is gone and customers had to migrate away).
  • Verdict: Confirmed dead as a standalone product. If considering GitKraken DevEx as the successor, that’s now a different (broader devtools) buying decision, not a documentation-tool decision — out of scope here.

8. Sourcegraph / Cody

  • 2026 status: enterprise-only. Free and Pro tiers shut down on 2025-07-23 (multiple 2026 review sites corroborate: aitoolsdevpro.com, weavai.app, dev.to/jovan_chan review). Cody now starts at $59/user/month; Sourcegraph Enterprise itself starts at ~$16K (deal floor) with AI-feature credits, cross-repo context, model controls. Source: sourcegraph.com/docs/cody
    • aggregator reviews above (cross-checked against multiple independent sources, consistent on the July 2025 free/pro shutdown date). by up to 10 repos simultaneously for cross-service architecture questions; bring-your-own-model (Claude Opus/Sonnet, GPT-4o, self-hosted).
  • Verdict: Priced and packaged for large orgs paying per-human-seat ($59/user/mo) — economically backwards for a 2-human team; the value proposition (cross-repo search + context for many human engineers) doesn’t match a 64-agent/2-human topology where the constraint isn’t “human search productivity” but “agent context correctness.” Not a fit; also the free tier that might have been worth a trial no longer exists.

9. Does auto-generated documentation actually stay fresh, or rot?

  • Primary academic evidence — “Context Rot in AI-Assisted Software Development: Repurposing Documentation Consistency for AI Configuration Artifacts,” Christoph Treude (Singapore Management University) & Sebastian Baltes, arXiv 2606.09090, dated 2026-06-09 (i.e., ~6 weeks before this research, directly on-topic for AGENTS.md/CLAUDE.md-style config files):
    • Analyzed 356 repositories; found 230 stale code-element references across 82 repos (23.0% of repos, 95% CI 18.8–27.2%) had at least one stale reference in their AI config artifacts (AGENTS.md-style files).
    • Of 29,454 candidate elements extracted, 18,048 verified as genuine references; of those, only 230 (1.27%) were stale — so the file-level rot rate (23% of repos affected) is the actionable number, not the low-looking 1.27% per-reference rate.
    • Manual validation of 50 sampled stale refs: 64% genuine referential rot (deleted deps, renamed/removed functions, deleted scripts, changed method signatures), 24% false positives, 12% ambiguous.
    • Their fix recommendation is exactly the freight-vs-markdown answer: treat config/doc files as code and review them in the same PR/refactor that renames or deletes the things they reference — i.e., process discipline (CI lint + PR review), not a new tool.
  • Secondary evidence, same conclusion, different angle: general LLM-doc literature (arxiv 2606.04397 on cross-file dependency chains for LLM-generated docs; smcmaster.com “Impressive But Wrong” blog) converges on: vague/outdated/hallucinated LLM-authored comments actively mislead more than no comment at all, because reviewers skim past prose that looks authoritative.
  • Conclusion on rot: auto-generated docs (DeepWiki-style, no human gate) demonstrably rot AND hallucinate — worse, they hallucinate confidently, which is more dangerous than staleness alone. Docs that are markdown-in-git, co-reviewed in the same PR as the code change (Swimm’s Auto-sync model, or plain CI lint), have a measured ~23%-of-repos rot rate even with that discipline — so the honest bar is “expect ~1-in-4 files to drift eventually regardless of tooling,” and the fix is cheap (grep/lint for renamed symbols in PR CI), not a paid product.

10. AGENTS.md — now a governed cross-tool standard

  • Released by OpenAI, August 2025. Adopted by 60,000+ repos as of the Dec 2025 announcement. On 2025-12-09, placed under the newly formed Agentic AI Foundation (AAIF), a Linux Foundation project, alongside Anthropic’s MCP and Block’s goose. Platinum members: AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, OpenAI. Sources: linuxfoundation.org press release, techcrunch.com/2025/12/09, openai.com/index/agentic-ai-foundation.
  • Read natively by: Claude Code (via AGENTS.md support, though Claude Code’s own native file is CLAUDE.md — the two co-exist; common practice by 2026 is AGENTS.md as the file, CLAUDE.md as a symlink to it, “one file on disk, two filenames, zero drift” per multiple 2026 guides), OpenAI Codex CLI, Cursor, Aider, Devin, Copilot, Gemini CLI, Windsurf, Amazon Q, Sourcegraph Amp, Google Jules.
  • Format is deliberately dumb: plain Markdown, no required fields, no YAML frontmatter, no special syntax — this is the “freight” question answered directly by the standard’s own design: the industry converged on plain git-committed markdown, parsed as prose as the interoperable substrate, not a proprietary format requiring a vendor tool to author or read it.

11. Claude Code Skills / SKILL.md — now a portable open standard too

  • Anthropic published Agent Skills as an open standard on 2025-12-18 (agentskills.io, spec + SDK). Folder-based Markdown modules with YAML frontmatter (name/description ~100 tokens, scanned at session start; full body <5,000 tokens, loaded on match — “progressive disclosure”). Adopted by Microsoft, OpenAI, Atlassian, Figma, Cursor, GitHub per launch coverage (opentools.ai, unite.ai, siliconangle.com, thenewstack.io — all 2025-12-18/19). Source: anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills.
  • Confirms the same pattern as AGENTS.md: the emergent “knowledge medium” for agents in 2025-2026 is markdown files in git, organized by convention (SKILL.md, AGENTS.md, CLAUDE.md), not a hosted/rendered knowledge-base product. Both standards were open-sourced specifically to prevent vendor lock-in around what is, mechanically, just structured prose in version control.
  • Independent pattern corroborating the same conclusion from a different origin: Andrej Karpathy’s “LLM Wiki” gist (gist.github.com/karpathy/ 442a6bf555914893e9891c11519de94f, posted ~April 2026, 5,000+ stars) — “Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.” Three layers (raw → wiki → schema), git repo of markdown as the artifact, the LLM (not a SaaS product) does the summarizing/cross-referencing/linting. Multiple community forks exist, including Agent-Skills-compatible versions for Claude Code specifically (github.com/Astro-Han/karpathy-llm-wiki).

12. llms.txt — adoption is real but thin, and unproven

  • Adoption exists (some major publishers ship one) but is inconsistent; no major AI provider (OpenAI, Google, Anthropic, Meta) has publicly confirmed reading/using it in production as of Q1 2026. Source: aeo.press “State of llms.txt in 2026,” caseyrb.com crawl study.
  • Casey Burridge’s crawl (caseyrb.com/blog/state-of-llms-txt-adoption/): monitoring AI bot traffic over 90 days found only 84 requests targeting llms.txt out of 62,100 total AI-bot visits — 0.1% of crawler traffic. SE Ranking analysis of ~300,000 domains found no statistically significant correlation between having an llms.txt and AI-citation frequency.
  • Verdict: llms.txt is a public-web/SEO-adjacent convention for external crawlers discovering your site; it is irrelevant to an internal-agent context problem (our 64 agents read the repo directly via tools, they don’t crawl a public URL) — not applicable to our use case except as a pattern Mintlify auto-generates for docs sites, which we don’t need.

Bottom line for the 1.05M-LOC EOL Ruby/Java + greenfield TS monorepo, 2 humans + 64 agents

  1. Every “AI wiki that reads your code and writes docs” product tested by independent parties (DeepWiki, and by extension Komment’s identical architecture) has a documented, named-and-dated hallucination failure mode on real codebases — not edge cases, core claims (build systems, install methods, function purpose). For an EOL legacy estate where the whole point is capturing correct tribal knowledge before it’s lost, a tool that invents plausible-but-wrong facts with no human gate is actively dangerous, not merely low-value.
  2. The two vendors whose architecture is defensible (Swimm, Driver AI) both explicitly pair LLM prose with deterministic static analysis — i.e., their pitch is literally “pure-LLM summarization (DeepWiki-style) isn’t enough.” That validates skepticism of the wiki-tool category generally, but both are priced/staged for enterprise migration engagements or are an unaudited 18-month-old seed startup — neither is a low-friction adopt.
  3. The industry’s own converged answer to “how do agents consume codebase knowledge” in 2025-2026 is: plain markdown files in git, addressed by filename convention (AGENTS.md/CLAUDE.md, SKILL.md), read directly by the agent’s own tools — not rendered through a hosted product. Two independent standards bodies (AAIF/Linux Foundation for AGENTS.md, Anthropic’s open Agent Skills spec) converged on exactly this, specifically to avoid vendor lock-in on what is mechanically just version-controlled text.
  4. Rot is real even for git-native docs (~23% of repos show stale AI-config references per the June 2026 study) but the fix is process, not a product: review docs/config files in the same PR that renames/deletes what they reference, same as you’d review any other diff. This is materially cheaper and more auditable than paying for a wiki-generation product whose failure mode is confident fabrication.
  5. Recommendation: no doc-generation tool purchase. Use plain markdown-in-git (AGENTS.md/CLAUDE.md at repo root, SKILL.md-style modules per domain, ADRs for the “why,” a CI check that flags renamed/deleted symbols referenced in docs) for both the legacy estate and the greenfield monorepo. If a paid pilot is wanted, Driver AI (MCP-served compiler-derived facts for the Ruby/Java estate specifically) is the only one architecturally aligned with “agents need verified facts, not narrated prose” — but treat it as an unaudited-vendor pilot with an exit ramp, not an adoption.