Doklo reads an AI-built codebase and turns it into product docs people can read, review, and hand over. Every claim traced to the code it came from, and your agents query the same truth over MCP.
free to start · 60 trial credits when you connect a repo
"Oh. This looks great."
two prompts in, the app runs
"So… how do I actually use it?"
the question Doklo was founded on
That second question doesn't go away. It comes back when you hand the project to a client, when a teammate inherits the code, when an agent edits a feature nobody documented. Two thirds of AI-generated code is "almost right", and the missing third is rarely more code. It's context: what this thing is, who it's for, and why it behaves the way it does.
"almost right": Stack Overflow developer survey, 2025
Doklo structures a codebase into Doks: the smallest units of business meaning. Login. Checkout. Permissions. Each one answers six questions.
The intent, in product language. No jargon required to read it.
Roles from the code, not from someone's memory.
The steps a person takes and what the system does in between.
Limits, permissions, and policies that explain "why can't I…".
Acceptance criteria you can hand to QA as-is.
Source anchors, injected deterministically. Never guessed by the model.
Code→Dok→Hub→Understanding· Doks live in your git as JSON. Together with roles, terms and screens they form the Hub.
Point Doklo at your repo. Deterministic parsing maps routes, components, state and the import graph. No model involved at this layer.
Doklo drafts Doks from the code — one call per Dok, one credit per Dok, checks always free. Source anchors are attached deterministically, outside the model's reach.
Every Dok arrives as a draft. You read it, fix it, approve it. Only what a person marks active becomes the record. Nothing publishes itself.
People read Live Docs rendered from approved Doks. Agents query the same Hub over MCP. One context, two audiences, zero forked truths.
Live Docs are rendered from reviewed Doks: a help page a client can follow, and a growing catalog behind it. When code drifts, the docs know.
Sixteen templates ship in the engine. One is promised stable today; the rest are opt-in while they earn it. That honesty is the product.
Pay for your cart in three steps. Your items are held for 10 minutes while payment completes.
Docs you can't trust are worse than no docs. Four guarantees keep the understanding honest, so you can lean on it.
Source anchors are injected from the import graph, deterministically. The model never invents them: zero hallucinated anchors across full production runs.
Each Dok carries a logic_hash of the code beneath it. When the code moves, the Dok flags stale everywhere it appears: Studio, Live Docs, MCP. One doklo sync catches it up.
AI writes drafts. People make records. The stable pipeline only renders Doks a human explicitly approved, and your edits survive regeneration.
Doklo clones your source only while an analysis runs, stores none of it, and destroys the copy when the job ends. The Hub keeps only what you explicitly connect: that's what publishing, MCP and drift detection stand on.
Cursor and Claude Code query the Hub over MCP: status, freshness and source anchors included. Answers cite Doks, Doks cite code, and a stale README stops being the smartest thing in the room.
The engine is open source and complete. You pay only when Doklo keeps the understanding alive on its own.
no card · nothing to install
for freelancers & small teams · waitlist
coming later · viewers free
checks are free, you pay only for generation · run out of credits? plug your own API key instead of a paywall
Connect a repo, watch Doklo draft its Doks, and publish the help page you'd hand to a client.