Your codebase has history. Your agent should be able to remember it.
Ask "why is this here?", "what did we already try?" or "when did we decide this?" across code, commits, design notes and every previous agent session. docenta hands the next agent the small set of evidence it needs, with exact citations and a disclosed token budget, over MCP, instead of the whole repository.
On 17 September 2026, in the design record, on the numbers from the winbox rebuild: of the 13,176 seconds the serial index apply spent, about 13,150 were 13,154 calls into the sparse encoder, so the lexical index of 7.5 million representations was hostage to one model lane. That reversed the decision of 4 September (commit b4f63231), which had put a measured gate on the encode inside the index step after one 2,675-unit batch took 86 minutes on an Intel laptop. The split landed the same day as commit ebffa66b: the index step never encodes, and a new sparse step back-fills.
Real, from the maker's own corpus, answered today: the design record, the git history and the agent session that made the decision are one collection over MCP, and the docent answers from all of it. This is how the maker's own coding agents work; the grep comparison below is the measurement.
Try the questions on git itself.
Your repositories are private, so the exhibit above is the maker's own. To let you test the claim on code you know, we sealed a public collection of the Git project: git/git at the sealed commit, all 82,305 commits since 2005 rendered as correspondence, and the git mailing list from lore.kernel.org, 552,249 messages. Below, four questions asked of it on 18 September 2026, answered the way the door answers over MCP. The door at git.docenta.ai opens shortly; its card, when it opens, is the one your assistant reads.
Because nobody liked the first two. On 22 July 2013 Junio C Hamano sent the third round of the series that had introduced the option as --cas and then --lockref, renamed: "Nobody liked 'cas' as it was too technical, many disliked 'lockref' because 'lock' sounded as if push by others were excluded by it while in fact this is to fail us." The name chosen says it is a kind of --force, and "with-lease" says the forcing holds only while the lease you took on the ref, the state you observed before you rebased, is unbroken. The same letter warns that the default lease, the remote-tracking branch, had been shown fragile in the discussion and is documented as experimental.
Real, from the sealed git collection, asked through the same engine the door serves. Three more, asked the same day:
Why does git rerere exist? Since 29 January 2006: Junio C Hamano's commit 8389b52b, "reuse recorded resolve", for a workflow of long-lived topic branches where the same conflict is resolved over and over; the list saw it the same day as [PATCH/RFC]. In July 2006 Shawn Pearce proposed a second tool so that git merge could reuse resolutions on its own, and in July 2007 a patch to enable rerere by default drew a thread of nine replies.
Which thread settled the SHA-256 transition? The plan was argued as "RFC: Another proposed hash function transition plan" through 2017, v3 in September; the function was settled in July 2018, when the cover letter "document that NewHash is now SHA-256" quoted Linus Torvalds on the libgit2 argument that the system's crypto libraries already carry it: "I think this is probably the single strongest argument for sha256. It's just there." The base SHA-256 implementation landed in November 2018.
When did git switch and git restore appear, and what did they refuse? Duy Nguyen's RFC v3 of 29 November 2018 introduced switch-branch and restore-files after two rounds of renames, listing what they refuse that checkout allowed: no invocation that does nothing, no implicit detach, no pathspec on switch. The branch nd/switch-and-restore cooked from April 2019 and merged on 9 July 2019; in May 2021 the list was still asking whether the two are experimental.
Each answer names the message or the commit it came from; the door returns the same citations to your assistant, and a question the collection cannot answer is answered "not in this corpus".
You remember the meaning, the date, the argument. Ask that.
Why
Why does this guard or workaround exist? Show the bug, incident, commit or session that introduced it.
Why was this dependency chosen over the other one?
Which incident or benchmark caused this retry, cache or timeout behavior?
What we tried
What did we try before this design, and why did we reject it?
Find previous attempts to fix this bug and say why each failed, with citations.
What did the previous agent already investigate about this failing test?
When
When did we decide to switch the importer to streaming, and why?
Which commit changed this schema or API, and what discussion explains it?
What is the last known-good commit before this regression, and what changed right after?
Contradiction
Where do the docs, the code, git history and agent sessions disagree about intended behavior?
Which TODO or comment is stale against later code and decisions?
Evidence and gaps
What context should a fresh coding agent read before modifying this subsystem?
Which design note states the invariant this function is supposed to preserve?
What can the current corpus not explain about this code path?
On the Git project, public
Why is it called --force-with-lease, and what was wrong with the other names?
Why does git rerere exist, and who asked for it?
Which thread settled the SHA-256 transition plan?
When did git switch and git restore appear, and what did they refuse that checkout allowed?
Code and git.
Every repository on the machine with its whole history: the commit that introduced the line, the files that change together, the message that explained it. C, C++, Go, Java, JavaScript, Kotlin, Python, Rust and TypeScript parsed by structure, identifiers as terms.
Notes and docs.
Design notes, incident write-ups, READMEs and architecture records, cited to the line, so the invariant a function protects has a source.
Every agent session.
Claude Code, Codex and ChatGPT sessions rendered and indexed as they happen, so what the last agent learned is one search away for the next, and a compaction loses nothing.
The same four-part contract, every time.
Exact source.
The file, the page, the message, the span. Not "sources": the line you can open and read.
Order and version.
When it was said, which version it was, what replaced what. A three-event timeline beats another paragraph.
Conflicts.
Two sources that disagree are shown side by side, never smoothed into one confident sentence.
Gaps.
"Not established by this collection" is an answer, printed as one. It is how you know the rest is solid.
The evidence path.
What was searched, which lanes answered, what was dropped and why, how many tokens the agent was handed. Inspectable, every time.
Boundedness is the feature.
Most AI marketing hides uncertainty. docenta prints it. The lines below are not disclaimers; they are what a trustworthy answer looks like.
A session explains; it can be stale. An agent's reasoning from February is cited with its date; a later commit that changed the code outranks it, and the answer says which is newer.
grep is not replaced, it is measured. On the maker's judged sets docenta's top-10 found 114 of 227 targets to ripgrep's 44, reading about 20 KB to the truth instead of 2.8 MB. The numbers are printed, the arm is committed.
Absent rationale is reported absent. "No design record in the corpus explains this" is an answer, and a useful one before the next rewrite.
The budget is on every answer. How many tokens the agent was handed, and what was dropped to stay inside them.
Back to the collections · also for contractors, lawyers, insurance, home, mail, developers, researchers · public collections, live: Judaism, Christianity, Islam, faith, sleep · in preparation: diabetes · the handbook
Ask the sensitive question without first uploading the sensitive file.
Private source trees, incident notes and agent transcripts stay on your machine: ingest, index, embeddings and answers all run there. Originals are never copied; only searchable derivatives are stored. Your agent gets a small evidence pack, a few kilobytes with the receipts, not your archive. Zero uploads, zero telemetry, and the token budget is printed on every answer.
It gets it over MCP, the connector standard Claude, ChatGPT and most coding assistants speak, from a server that runs on your own machine: the agent talks to a local process, not to a URL in someone's cloud. The only remote MCP addresses on this site are the public doors, sealed collections hosted on purpose; your own collection never has one. Deterministic layers, no learned model in the retrieval loop, bit-stable re-runs.
One address, one sign-in, every answer cited to the commit or the message.
The sealed Git collection above is served as an MCP server at https://git.docenta.ai/mcp: 653,394 documents, the git/git tree at the seal, every commit since 2005 and the whole mailing list. Your own repositories never have an address like it; this door is where you test the claim on code you know.
Register once
Use the form at the bottom of this page and say "developers". You confirm your address with one click, and that registration is what opens the collection for you. There is no key to copy, ever: the address below is not a secret, and signing in is what makes it yours.
Add the address to your assistant
https://git.docenta.ai/mcp
Claude Code: one line, then type /mcp and press Enter on docenta-git.
claude mcp add --transport http docenta-git https://git.docenta.ai/mcp
Claude (web and desktop): Settings, Connectors, Add custom connector, paste the address, Connect. ChatGPT: Settings, Connectors, Create, paste the address, choose OAuth.
Allow it, once
Your assistant opens a docenta page by itself. It asks who you are (your address and six digits from your mail; nothing at all if you come from your start page) and shows one sentence: which assistant, which collection. Press Allow. You can disconnect any assistant later on your start page.
Ask, and read the citation
Every answer names the document and the page it came from. The first question after a quiet spell can take half a minute while the collection wakes up; your assistant is told so and asks again by itself.
The door never sees your files, and cannot be changed by anyone, us included, without a new seal and a new date on this page. It knows that a registered person asked, so each person gets a fair share of it. It keeps nothing you ask unless you choose to, with two boxes that start unticked when you connect: "tell me when a question can be answered later", which keeps only the questions the collection could not answer, and "keep my questions, without my name", which keeps the words for 180 days with no name, address or time of day beside them. You can change both on your start page.
What is one question your files should be able to answer today?
docenta is a commercial product of SKY, LLC, now in private preview. Tell us the question; we will tell you when the door opens, and the questions we hear shape what we build.
No account and no newsletter: you confirm your address by mail, then we write back when the door opens. Prefer mail? developers@docenta.ai