Your team already
worked this out.
Nobody can find it.
DocBrain remembers what your engineers figure out — in incidents, pull requests and threads — and warns whoever touches that code next, human or AI agent, before the change lands.
It records what nobody wrote down
The fix found at 2am. The decision from the review. Read in place, read-only, from the systems your team already uses.
It reaches the next person in time
Not a wiki someone has to think to search. It arrives in the editor and in the agent, before the change — or it may as well not exist.
And nobody can quietly rewrite it
Every claim cites its source, expires when reality moves, and exports as proof a stranger can check offline.
Rewrite history. Type whatever you want her to have said.
This is the actual sentence sealed inside the evidence bundle at 02:47. Change a word, change all of it, put your own name on it. Then seal it and check. The verifier runs on the real bytes, in your browser, with no network.
Or try the attacks a real forger would
One sentence, kept for eight months, and it paid for itself in three seconds.
Priya never wrote a document. She never filed a ticket. She said one true thing in a channel at 02:41 and went back to bed.
Captured it where it happened
Read the incident thread in place and kept the one sentence that mattered, cited to the message and the minute it was said. Nobody had to remember to write anything down.
Reached the next person in time
Eight months later: a different team, an engineer who had never met her, and an agent that stopped the change and named its source — before it shipped, not after the next incident.
Told the truth about itself
When the chart changed, the answer marked itself stale instead of staying confidently wrong. And the record is signed, so nobody can quietly rewrite what was decided.
That is the whole product. Everything below is how you run it, what it costs, and where we lose to the alternatives.
What it costs to run, and who has to look at it.
"Self-host in 5 minutes" and "run this for 400 engineers" are not the same sentence. Here is the second one.
Four services, one origin
PostgreSQL, OpenSearch, Redis and the server behind a single origin. Docker Compose for a trial, a Helm chart for production. No message broker, no separate vector database to operate.
Modest until you train
The chart requests 500m CPU and 2Gi for Postgres, a few hundred megabytes each for server and web. The only heavy component is the optional trainer — 2Gi, capped at 8Gi. Leave learning off and you never deploy it.
You choose, so you control it
Inference runs on your provider or fully locally through Ollama. We never sit between you and the model — which is why agents are unlimited: the marginal query is your cost, not a seat we resell.
SSO on day one
Generic OIDC, GitHub, GitLab, Azure AD and ADFS, with role mapping from your directory groups. Four-tier RBAC and per-space isolation.
Audit and retention
Admin actions and access decisions are written to audit tables. Retention is set per source with scheduled purges, so Slack can expire in 30 days while Confluence is kept indefinitely.
Somebody does have to review
Drafts and repair proposals land in a review queue, and that is real work. It is bounded by your own activity rather than by us, and it batches.
We don't charge for access. We charge for the memory we keep true.
Never per person. Add anyone, add any number of agents, and the price doesn't move. Band pricing sized by the connected user base, not by seats — unlimited people, unlimited AI agents, unlimited questions, captures and generated docs inside the band.
| Band | Users in connected scope | Annual, flat |
|---|---|---|
| Community | up to 25 | Free, forever |
| Growth | 26 – 100 | $10,000 |
| Scale | 101 – 300 | $25,000 |
| Enterprise | 301 – 750 | from $60,000 |
| Enterprise+ | 751 – 2,000 | Contact us |
| Global | 2,001 and up | Contact us |
Annual, paid up front. "Users in connected scope" means the population whose sources are plugged in — one org, a division, or the whole company. You choose. Growing the account is a scope conversation, never a per-head negotiation with your finance team.
The entire product
Genuinely — not a crippled tier. Every connector, ingestion and search, answers with citations, document generation, IDE and agent integration, CLI, dashboard, Slack, quality scoring, freshness tracking, contradiction detection, review workflows, governance, access controls, SSO, audit logging, Autopilot, Docker and Kubernetes. Unlimited users, unlimited agents.
A relationship, not a feature gate
We don't hold features back to create a reason to pay. What you get is us, on the hook and reachable: the engineers who wrote the code, for deployment, tuning and upgrades — not a support queue — plus priority input on the roadmap and the connectors you need.
What it's worth, with the arithmetic shown.
Your numbers, not ours. Where evidence exists we cite it. Where it doesn't, we say so.
Minutes seeded from Stack Overflow's 2022 developer survey (61% spend 30+ minutes a day searching, 47% the same answering). They sell a competing product, so only raw responses are used. Suggest 20–30% removal for year one — measured in a pilot, not assumed.
—
What this model doesn't know: your actual removal rate. No vendor's does, including those quoting one. Both inputs are hypotheses to test in a pilot.
Including the rows where we lose.
If you're evaluating three tools, you deserve this from all of them. Two of these rows go against us and they're still here.
| Capability | DocBrain | Open-source search | Enterprise AI search |
|---|---|---|---|
| Captures what was never written down | Yes, at the source | No, retrieval only | No, indexes what exists |
| Citation granularity | Per claim | Per document | Per document |
| Declares what it cannot answer | Refuses, and logs the gap | Confidence score | Confidence score |
| Reads live systems at answer time | Yes | Index only | Index only |
| Retrieval tuned to your corpus | Optional, on your infra | Fixed model | Fixed model |
| Writes documentation unprompted | Autopilot, human-approved | No | No |
| Runs on your own AI provider | 14, local included | Yes | Vendor-hosted |
| Licence | Public source MIT; server closed, images BSL | MIT, more permissive than ours | Proprietary |
| Published accuracy benchmark | Not yet | Yes, with methodology | Varies |
| Named public customers | None yet | Yes | Yes |
If a permissive licence and a published benchmark matter more to you than capturing knowledge nobody recorded, an open-source search tool is the better buy — and we'd rather you learned that here than three months into a pilot.
A product built on refusing to overclaim shouldn't overclaim about itself.
Before you spend a week on a pilot, here is what we'd want to know if we were you.
The server is closed source.
Everything runs inside your network, server included. Everything in our public repository is MIT and auditable — the CLI, the MCP server, the Helm charts, configuration and docs. The server ships as a container you run but cannot read, and it is the component that touches your data. We're not naming a date for opening it until we're certain we'll hit it.
We have no accuracy benchmark to show you.
Grounding is measured internally and every model version is gated on it, but we won't put a percentage on a marketing page until it has been measured across real customer corpora and we can publish how we measured it.
Learning is optional, and off until you turn it on.
It needs real feedback volume and some compute before it earns its keep. Everything else works without it, and most teams should run that way first.
Self-hosting isn't unique to us.
There are open-source alternatives and several are genuinely good at retrieval. One is more permissively licensed than we are. What they don't do is capture what was never written down. Judge us on that, not on deployment.
Bring us the version of this you disagree with: hello@docbrainapi.com
Answered before you have to send the email.
We already have Confluence, Slack, Jira and GitHub. Why this?
Those are places to store knowledge somebody remembered to write down. DocBrain captures what never gets stored — the reasoning in a ticket, the fix in an incident thread, the decision made in a meeting — connects it across all of them and keeps it accurate. You're not replacing Confluence. You're capturing what Confluence never sees.
Does anything leave our network? Do you train on our data?
No, and no. DocBrain runs entirely in your infrastructure and reads your systems without ever writing to them. Point it at a local model and nothing crosses your boundary at all. If you switch on learning, that happens inside your infrastructure too, on your hardware, on your data. We never see it, because there is no "we" in the path.
What happens if we stop paying?
You keep running the version you have, in production, for as long as you like. You stop receiving updates, patches and support. Nothing is switched off, nothing phones home, and no data is held hostage, because it was never ours to hold.
How is this different from enterprise AI search?
Search retrieves documents that already exist. If the answer was never written down, better search returns nothing, faster. DocBrain's job starts one step earlier — capturing the answer at the moment someone works it out, so there's something to retrieve at all.
Knowledge that can be proven wrong.
Self-host it today, free, at any size. Or spend thirty minutes with the people who built it.