Detailed Analysis
A developer has open-sourced tela, a self-hostable documentation suite that treats AI agents as first-class users rather than an afterthought bolted onto an existing product. Built with Go and Postgres and released under the AGPL license, tela combines docs, presentations, and spreadsheets into a single wiki-style platform, but its distinguishing feature is a built-in MCP (Model Context Protocol) server embedded directly in the backend. This means Claude—whether accessed through the desktop app, Claude Code, or the standard connector—can search, read, author, and retrieve cited answers from an organization's internal knowledge base using scoped, per-tool permissions rather than a generic file-access bridge bolted on after the fact.
The more consequential capability is tela's ability to connect directly to a git repository or Jira instance and automatically generate and maintain living documentation of a codebase, with every factual claim traceable back to a specific commit or issue. Rather than producing documentation that quickly goes stale or hallucinated summaries divorced from ground truth, tela's design forces answers to be built from retrieved, cited sources. When a team member asks a question, the response is grounded in actual project history rather than a model's best guess. In effect, this turns Claude into a persistent, queryable expert on a team's own codebase, one that stays synchronized with reality because its answers are anchored to verifiable artifacts rather than a static training snapshot or an unmoored context window.
This project sits at the intersection of two trends reshaping how developers build with large language models: the rapid standardization around MCP as the connective tissue between AI assistants and enterprise data, and the growing recognition that grounding and citation are the difference between a useful AI tool and a liability. Since Anthropic introduced MCP in late 2023 and continued expanding it through 2024 and 2025, the protocol has become a de facto standard for exposing tools, files, and structured data to models like Claude in a permissioned, auditable way. Independent developers building MCP servers for niche and vertical use cases—like internal documentation—demonstrates that the protocol's value isn't confined to Anthropic's own integrations; it is spawning a broader ecosystem of interoperable, agent-ready infrastructure that any team can self-host without depending on a vendor's roadmap.
The emphasis on citations tied to commits and issues also reflects a broader industry shift away from treating LLMs as oracles and toward treating them as retrieval-augmented reasoning engines that must show their work. As organizations increasingly deploy Claude and similar models for internal knowledge work, the risk of confidently wrong answers becomes a bigger liability than slow answers. Tools like tela that hard-wire provenance into the documentation-generation pipeline point toward a future where enterprise AI assistants are judged not just on fluency but on auditability—an especially important property for regulated industries, open-source maintainers, and any team that needs to trust an AI-generated explanation of why code changed the way it did. The AGPL licensing choice and self-hosting model further signal a segment of developers who want Claude's agentic capabilities without ceding control of proprietary codebases to third-party SaaS platforms, a tension likely to persist as more of these agent-native tools proliferate.
Read original article →