AI tools that document a codebase in 2026: 17 compared, and what none of them does yet

October 2, 2026

Most of what a team needs to know about a system is locked in its source code. Developers can read it, slowly. Managers, business stakeholders, support staff and coding agents cannot, so they rely on documents that are missing, out of date, or written for someone else. A wave of AI tools now promises to fix that by turning a repository into documentation. On 1 October 2026 I reviewed seventeen of them, hosted and self-hosted, free and enterprise, and compared what each one reads, what it produces and who it is for. Developer wikis turn out to be a commodity. What is still missing is more interesting.

In brief

1. Developer wikis are a commodity. Three hosted products (DeepWiki, Google Code Wiki, Zread) and at least five active open-source projects turn a repository into a developer wiki with diagrams and chat.
2. The open-source bar is high. CodeWiki 2.0 does graph-based decomposition and component-level incremental updates. OpenDeepWiki ships an MCP server, an admin console and scheduled incremental updates on .NET 10 with Microsoft’s agent framework.
3. Audience-specific documents exist only in shallow form. Microsoft’s deep-wiki plugin writes onboarding guides for four audiences, Executive and PM among them, and a few small projects write business documents from code. All of them are prompt-only: no persistent index, no metrics, no upkeep.
4. Deep business-audience output exists only in enterprise legacy tools. Swimm and Kodesage extract business rules for analysts and product managers, but target COBOL, Oracle Forms and similar systems, and sell through sales teams.
5. Manager metrics and narrative documents live in separate products. CodeScene has hotspots, knowledge distribution and portfolio views with no generated documentation. The wiki tools have narrative with no git-derived metrics.
6. The real gaps are a maintained manager and business view for ordinary modern repositories, claims a reader can verify, with a visible line between what was read from the code and what a model interpreted, and a knowledge base for agents built as structured facts instead of a re-served wiki.

The market in five segments

Hosted repository wikis

DeepWiki, Cognition’s Devin Wiki, is the reference product in the category: a structured wiki with architecture diagrams and source-linked sections, an “Ask Devin” chat grounded in the wiki, free for public repositories and available for private ones through a Devin organization. Its MCP server has three tools, read_wiki_structure, read_wiki_contents and ask_question, and a .devin/wiki.json file steers page structure and adds repository notes. Zero setup, more than 50,000 popular repositories pre-indexed (the vendor’s figure), source links on claims and date-stamped re-indexing. It is closed source with no local or offline use and no choice of model; private use is tied to Devin’s usage-based billing; the free MCP server covers public repositories only; it serves developers only, and reviewers note the volume can overwhelm newcomers; large repositories can hit built-in page limits.

Google Code Wiki is a Gemini-generated wiki for public GitHub repositories, in public preview since November 2025: a wiki regenerated as the repository changes, diagrams and links into the code, and a Gemini chat that uses the wiki as context, free and with no account needed. Regeneration after changes addresses drift. It is public repositories only on the website; private-repository support, a Gemini CLI extension that runs locally, was still waitlisted in the latest source reviewed, from May 2026; Gemini only, no control over structure, developer audience only, a preview product, and no MCP server found in the sources reviewed.

Zread (Zhipu, Z.ai) is a hosted AI wiki and code-reading assistant with architecture overviews and Q&A in English and Chinese, an MCP server with search_doc, read_file and get_repo_structure, a CLI and an agent skill, and private-repository support according to a third-party listing. Multilingual output, and MCP tools that give agents both docs and file access. Proprietary, the MCP server needs a Z.ai API key, the English documentation is thin, pricing is not clearly published, and it serves developers only.

Open-source wiki generators

deepwiki-open (AsyncFuncAI) is the most-starred open-source clone: 18.1k stars, MIT, last commit September 2026. A self-hosted web app (Python backend, Next.js frontend) that writes a wiki with Mermaid diagrams for GitHub, GitLab and Bitbucket repositories, private ones through a personal access token, with RAG-based chat, a multi-turn “DeepResearch” mode, Gemini, OpenAI, OpenRouter and local Ollama models, and “Codemap” guided tours. A large community and a simple Docker setup, on your own keys and local models. But the English README is now a stub pointing to Grok Wiki, the author’s newer desktop app that runs on local CLI agents, so the original project’s future is unclear; retrieval is embedding-based with a cache and no documented incremental update; no MCP server is documented; developers only. Grok Wiki itself adds source-cited wikis, “outcome-led” formats (mental models, hidden quirks, debugging maps, tech-reader briefs) and slide previews, and is young: 46 stars at last check.

OpenDeepWiki (AIDotNet) is an open-source repository knowledge base: 3.6k stars, MIT, last commit 30 September 2026. It imports from Git URLs, ZIP archives or local directories; generates overviews, wiki catalogs, page content, translations and mind maps; offers repository-scoped MCP endpoints, built-in chat and an embeddable chat API; has an admin console for repositories, users, roles, API keys and AI providers and models; runs background workers with scheduled incremental updates; and stores in SQLite or PostgreSQL. The stack is ASP.NET Core on .NET 10 with Microsoft.Agents.AI, a Next.js frontend and SQLite or PostgreSQL, and it has the most complete self-hosted feature set of the open-source group. It is a server application rather than a CLI, so heavier to run; developers only, with no manager or business outputs; enterprise support is paid; and the community and its chat channels (Feishu, QQ, WeChat) lean Chinese-language.

CodeWiki (FSoft-AI4Code) is a research-backed generator published at ACL 2026: 1.7k stars, MIT, last commit 30 September 2026. It builds a dependency graph, clusters it into a module tree, and has agents write one page per module, for Python, Java, JavaScript, TypeScript, C, C++, C#, Kotlin, PHP, Ruby, Scala and Rust. Build files, CI workflows, Dockerfiles and schemas are documented as graph nodes; --update patches only affected modules; an MCP mode hands its analysis tools to an IDE agent; there is an HTML viewer for GitHub Pages, an output-language option and custom instructions. It has the strongest published method for large repositories (tested on a 1.4M-line codebase), reports an incremental update at $0.34 against $21.48 for a full build on one repository, runs on API keys or on a Claude or Codex subscription, and ships its own benchmark, CodeWikiBench. The benchmark results are self-reported, it serves developers only, it is CLI only, the MCP mode exposes analysis tools rather than a documentation knowledge base, and source citations in pages are not described in the README.

Litho and Terrain (sopaco) are Rust tools; Litho has 3.1k stars and Terrain, its successor, 249, both MIT, Terrain last committed on 29 September 2026. Litho writes C4-model architecture docs (context, container, component, code) from a multi-agent pipeline, mounts external knowledge (PDF, Markdown, SQL) and documents database schemas with ERDs. Terrain produces two tracks from one codebase, narrative C4 docs for humans and a compact context file for agents, plus a business glossary, with incremental regeneration and a freshness score per asset, as a desktop app and CLI on OpenAI-compatible endpoints, Ollama or LM Studio. The explicit human-versus-agent split and the freshness scoring are unusual and useful, and the core scanning works offline with no model call from a single binary. Terrain is new and small, its installers are unsigned, agent access is through files and a CLI rather than an MCP server, and there are no manager or business documents beyond the glossary.

PocketFlow Tutorial-Codebase-Knowledge (Code2Tutorial) turns a repository into a beginner tutorial: 12.7k stars, MIT, last commit May 2026. It identifies core abstractions and writes chapter-style tutorials from GitHub repositories or local directories, with include and exclude filters, an output-language option and response caching, on Gemini by default with other providers and Ollama supported. Very readable for newcomers (a third-party comparison found it easier to follow than DeepWiki) and a small, simple codebase to learn from. It is a tutorial rather than reference documentation, about ten abstractions by default, with no incremental updates, no MCP and no site generator, and development has slowed.

Microsoft’s deep-wiki plugin is a prompt-and-skills plugin for GitHub Copilot CLI, MIT, in a parent repository with 3.1k stars. /deep-wiki:generate writes a full wiki with Mermaid diagrams and citations; /deep-wiki:onboard writes four audience guides, Contributor, Staff Engineer, Executive and PM; /deep-wiki:agents and /deep-wiki:llms write AGENTS.md and llms.txt; /deep-wiki:build and /deep-wiki:deploy package a VitePress site and a GitHub Pages workflow; and there are changelog, deep-research, Q&A and Azure DevOps wiki conversion commands. It is the closest existing match to “several audiences, a static site and agent files” in one tool, free, and already ported to other agents by the community. But it is prompts only, with no index, graph or stored facts, so quality depends on the host agent; the Executive and PM output are onboarding guides, not maintained documents; there are no metrics from git history, no incremental engine and no MCP knowledge base; and it is built for Copilot CLI.

GitNexus is a code knowledge graph for agents: 47.7k stars, last commit 1 October 2026. It indexes a repository into a graph of dependencies, call chains, clusters and execution flows, exposes MCP tools for impact analysis and context retrieval, generates module pages and an overview from the graph with gitnexus wiki, has a web UI with a graph explorer and chat, and writes AGENTS.md and CLAUDE.md. Strong structural analysis that runs locally, with wiki pages that cross-reference the graph. Its licence is PolyForm Noncommercial, so commercial use needs a paid licence; documentation is a side feature for a developer and agent audience; and full incremental indexing is still on the roadmap.

Documentation platforms with agents

Mintlify is a hosted documentation platform: a docs-as-code site with a web editor and API playground, an agent that drafts and updates pages through pull requests, an Autopilot that watches repositories for user-facing changes and proposes doc updates, and an MCP server and llms.txt for the published docs. Polished published output and good drift handling for product documentation. It is aimed at external product and API docs rather than at understanding an internal codebase; pricing sources conflict (a free Starter tier is confirmed, third parties cite $300 a month for the agent tier, and Mintlify’s own docs place Autopilot on Custom plans); hosted and proprietary.

Dosu is a hosted knowledge layer for engineering teams and agents. It generates and maintains docs from code, pull requests, issues and chat threads, offers an MCP server that supplies context to agents and flags documentation gaps, syncs with Confluence, Notion, Coda, Slack and Teams, and does issue triage and public Q&A spaces for open-source projects. It uses sources beyond code, which is where the “why” lives, and is free for public repositories. It is proprietary SaaS with credit-based pricing (a third-party listing shows Pro from $16 a month, not verified); its claim that it adapts content for non-technical readers is the vendor’s; and there is no static site or audience-specific document set.

DocuWriter.ai is a paid SaaS generator: a full documentation tree for one or several repositories, API references, UML diagrams and Swagger or OpenAPI output, an Autopilot agent that keeps docs in sync, and claimed support for modern and legacy languages including COBOL and ABAP. Broad output types and multi-repository documentation; proprietary, pricing not verified here, developers only, and little independent evidence of quality.

Enterprise application understanding

Swimm is an enterprise “application understanding” platform, now focused on legacy and mainframe modernization: deterministic static analysis with AI explanation on top, business-rule extraction organized by business function, navigation from business capability to process to code step, rules turned into modernization specs and user stories, MCP access for agents, and on-premise and air-gapped deployment. It is the most serious business-audience product found, built for analysts and product managers, and its deterministic analysis reduces invented rules: Swimm’s own test reported full coverage against lower coverage from an LLM agent (vendor-run, so treat with caution). It targets COBOL, JCL, PL/I and similar, with little help for C#, TypeScript or Python repositories; sales-led pricing; closed source; out of reach for individuals and small teams.

Driver is an enterprise “context layer” that began as a documentation tool: pre-computed, symbol-complete documentation for every file, architecture maps and history guides, served to coding agents through MCP, for GitHub, GitLab and Bitbucket, with SSO, SCIM and cross-repository context. Exhaustive coverage computed ahead of time, across repositories. Contact-sales pricing with no free tier, enterprise focus, closed source, developer and agent audience only.

Kodesage is an on-premise platform for legacy systems that raised a $6.6M seed round in June 2026. It ingests code, existing docs, databases and ticket systems into one knowledge layer, produces living documentation readable by people and agents, does code conversion, test generation and a production-support assistant, and runs on-premise, in a private cloud or air-gapped. It combines code with non-code sources, explicitly serves business analysts and product managers, and has a strong data-residency story. Its focus is COBOL, PL/SQL, Oracle Forms, PowerBuilder and RPG; enterprise sales with no public pricing; closed source.

The manager view next door

CodeScene is behavioral code analysis, not a documentation generator, but it owns the manager audience today: hotspots from change frequency and code health, knowledge distribution and bus-factor maps, team-to-code alignment and delivery metrics, a portfolio overview and PDF reports. Mature metrics computed from git history, free for open-source projects, with an on-premise option. No narrative documentation, €18 to €27 per active author per month, and a dense interface with a learning curve.

Smaller and adjacent tools

Side by side

Counting the seventeen tools that offer each output outright says where the effort has gone: nearly every tool writes developer documentation and keeps it updated, half serve an agent, and the manager and business audiences are served by one and two tools, both built for other reasons.

Legend for both tables: yes, part (partial), no, and a question mark where the sources reviewed do not say.

Access, deployment and cost:

Outputs and audiences:

Read together, the two tables sort the field into a grid. The tools that run on your own machine and code are the open-source ones and the enterprise on-premise platforms; the hosted products are the ones with the least to configure and the least control.

What none of them does yet

1. A maintained manager document. No tool combines git-derived metrics (hotspots, ownership concentration, dependency age, test presence) with a written system summary and a “what changed” digest. CodeScene has the numbers; the wikis have the prose.
2. Business documentation for modern stacks. Capability catalogs and business-rule extraction exist only for mainframe and Oracle-era systems, behind enterprise sales. Nothing open source does this for C#, TypeScript or Python with any rigor.
3. Audience documents that stay current. The Executive and PM guides from Microsoft’s plugin are one-off prompt outputs. No tool regenerates non-developer documents when the code changes.
4. Observed versus inferred. Source links are inconsistent across tools, and none labels which statements are read directly from the code and which are the model’s interpretation. This matters most for business readers, who cannot check the code themselves.
5. A knowledge base shaped for agents. MCP servers today either re-serve the human wiki (DeepWiki’s three tools) or expose a raw code graph (GitNexus, CodeWiki’s MCP mode). Structured facts with stable IDs, citations and freshness are rare; Terrain comes closest, through files.
6. A structured place for human context. DeepWiki has repo_notes, CodeWiki has --instructions, and Litho mounts external documents. None has a defined input for business goals, owners and glossary corrections that feeds non-developer output.
7. Cost preview. Only CodeWiki publishes cost figures, and none of the tools reviewed advertises an estimate before a run.
8. System-level documentation across repositories. Driver offers cross-repository context for enterprises. The open-source tools work one repository at a time.
9. Quality measurement for non-developer docs. CodeWikiBench scores developer documentation. Nothing evaluates manager or business documents.

Two lessons follow for anyone building in this space. Audience-flavoured prose is one prompt away, as Microsoft’s plugin shows, so the defensible part of a new tool is what a prompt pack cannot copy: a persistent index, citations, a visible line between observed and inferred, and metrics computed from history. And Swimm’s argument that model-only extraction both misses and invents business rules is self-interested but plausible, so business output needs static analysis underneath it and explicit labels on top.

What is already well served

A developer wiki with diagrams and chat; private and local repositories on your own model keys (OpenDeepWiki, CodeWiki, Terrain); incremental updates (CodeWiki, OpenDeepWiki, Terrain); MCP access to documentation; static site output (CodeWiki’s HTML viewer, deep-wiki’s VitePress build); multilingual output. Competing on any of these alone is a large effort for little distinction.

Method and caveats

• Vendor documentation, GitHub READMEs and third-party reviews, read on 1 October 2026. Nothing was installed or run, so quality claims are the vendors’ or reviewers’ own.
• GitHub star counts and last-commit dates were read from GitHub on that day.
• Prices are quoted only where a source states them; where sources disagree, the article says so.
• Sources beyond the vendors’ own pages: a nine-tool roundup from June 2026 and a comparison of DeepWiki and Code2Tutorial.

I’m Amir Pournasserian. I build AI and platform systems for a living, maintain FluentCMS and YeSvelte, and write here about what I find along the way.