DocBrain captures knowledge from PRs, Slack threads, CI pipelines, and your IDE the moment it's created, connects it into a living graph, and scores the quality. Then one command — docbrain generate — turns it into a grounded runbook, postmortem, or reference, right from your terminal or your CI pipeline.
The problem isn't missing docs. It's that knowledge gets created in one place and written down in another. If it gets written down at all.
A senior engineer explains a critical architecture decision in a thread. Two months later, nobody can find it.
Every merged PR contains decisions, caveats, and procedures. But PR descriptions are write-once, read-never.
Code has linters, tests, and review workflows. Documentation has nothing. No quality scores, no style enforcement, no SLAs.
How It Works
Knowledge captured at the source, quality-scored automatically, published through configurable review workflows.
Knowledge fragments auto-extracted from merged PRs, Slack conversations, CI deployments, and IDE annotations.
Three-layer quality pipeline: structural analysis, team-defined style rules, and LLM-assessed semantic scoring.
Fragments cluster by semantic similarity. When enough accumulate, DocBrain auto-composes them into documentation.
Space ownership, SLA policies, multi-stage review workflows, and breach detection.
Other tools index existing docs. DocBrain captures the knowledge that was never written down and turns it into quality documentation.
Result: documentation grounded in real engineering work, quality-scored and governed, without asking anyone to write docs.
Generate · The Moat
docbrain generate drafts a runbook, postmortem, or API reference grounded in your runbooks, incidents, tickets, and PRs — with per-claim provenance. A frontier model writes fluent, generic prose. generate writes what is true for you, or tells you what it can't answer. Run it from your terminal today, or wire it into CI so docs update the moment a PR merges.
stdout is the markdown (pipe-clean). Diagnostics go to stderr. Swap --source notes.md for --source-url to ground in a Confluence page, Jira issue, Slack thread, or GitHub PR.
Drop --out and the exit code becomes a gate: generate exits non-zero on error-severity quality violations, so a sub-bar doc fails the build. If a named source can't be fetched, the whole run hard-fails — never a doc from a partial set.
Reads your real PRs, runbooks, incidents, and threads through your connectors. Every section is attributed to the source it came from — reviewers verify, they don't trust.
When the knowledge isn't there, it emits needs_input — the open questions — instead of fabricating. You fix gaps on purpose, not mid-incident.
Secret/PII redaction, hostname scrub, injection-quarantine, and a 0–100 quality score run on every draft. The exit code is your CI signal — "good enough to publish" becomes machine-checkable.
Access Everywhere
One knowledge layer, five interfaces. Ask from Slack during an incident, from your IDE while coding, or from the terminal at 2am.
git log --oneline production..mainkubectl rollout undo deployment/payment-processorcurl -sf https://api.example.com/health/paymentsLive Intelligence
Most doc assistants answer from a snapshot. An index tells you what was true the last time it crawled, not what's happening this minute.
DocBrain reaches into your live systems at the moment you ask. “What's the status of PROJ‑412?” pulls the ticket from Jira as it stands right now, then combines it with your indexed runbooks into one cited answer. Old knowledge and live state, in a single response.
DocBrain can read your tools, never write to them. Any tool that could change a system is dropped at discovery. Enforced, not promised.
Connected per‑user over OAuth. Your answer is scoped to what you can see in the source. Tokens refresh on their own, and nobody sees data they shouldn't.
Jira, Confluence, and Slack out of the box, plus anything that speaks the Model Context Protocol. Add a tool, and DocBrain starts answering from it.
Write operations
OAuth scope
At answer time
“Is the payments outage from yesterday resolved, and what was the fix?”
PROJ‑412 (“Payments 502s after deploy”) moved to Done at 09:14 UTC. The fix was a rollback of payment-processor plus an idempotency-key guard, per the linked runbook.
One answer. Indexed history reconciled with live source-system state.
True. Cursor and Claude Code can hit your tools live too. The difference is what happens after the answer. They forget it. DocBrain keeps it, connects it, and gets better every time someone asks.
An IDE asks your tools a question. DocBrain turns every question your org has ever asked into a system that gets smarter.
4-Layer Memory System
Four memory layers work together so every question makes the system sharper. The 100th query about Kubernetes is faster and better-answered than the first.
Cache hit rate
Avg response
Binary (Rust)
Multi-turn conversation context. Resolves "it", "that service", "the same thing."
Every Q&A ever asked. Semantic caching, feedback learning. Gets cheaper as it learns.
Entity relationships via graph traversal. "Who owns the payment service?" gets resolved, not searched.
Adapts retrieval from feedback. Discovers: "deployment questions → search DevOps first."
Features
First-class knowledge units with provenance. Auto-captured from PRs, conversations, and CI pipelines.
Merged PRs and deployments auto-analyzed by LLM. Decisions, caveats, procedures extracted as fragments.
Three-layer scoring: structural, style rules, and LLM-assessed semantic quality. Composite score 0-100.
Hybrid search, intent classification, cited answers with freshness. Low confidence asks questions instead of hallucinating.
Entity relationships with BFS/DFS traversal. Blast radius analysis, expertise routing.
Clusters unanswered questions, detects gaps, auto-drafts missing documentation.
docbrain generate drafts a runbook, postmortem, or reference grounded in your own runbooks, incidents, tickets, and PRs — with per-claim provenance, or an honest list of what it can't answer instead of confident fiction. Same redaction, injection-quarantine, and quality gates as every DocBrain doc. Returns the markdown — you decide what to publish. Wire it into CI to update docs on merge.
Anthropic, OpenAI, Bedrock, Ollama (local), Gemini, Azure, and 8 more. Swap via config.
GitHub/GitLab/OIDC SSO. Four roles. Space-level isolation. API keys Argon2-hashed.
Mirrors real Confluence / Slack / GitHub / Jira permissions at query time. Restricted page in source = filtered out for users who can't read it. Three modes, side-channel-safe, audit-logged.
25MB binary. <500ms cold start. Under 100MB memory. Self-hosted, your data stays yours.
Plug in any knowledge source in any language. 3 endpoints, DocBrain handles the rest.
10 tools for Claude Code, Cursor, and any MCP editor. Capture decisions at commit time.
Pull live Jira tickets, GitHub PR state, or any MCP-enabled system at answer time. Per-user OAuth with transparent refresh. Answers cite live data, not stale chunks. See how →
Confluence, Slack, Teams, GitHub, GitLab, Jira, PagerDuty, OpsGenie, Zendesk, and more.
Shift-Left In Practice
Knowledge captured where the work happens. Not after. Not in a doc sprint. At the source, in real-time.
9:15 AM · PR Merged
DocBrain's CI capture extracts 3 knowledge fragments: a decision about retry logic, a caveat about idempotency keys, and a procedure for manual refund overrides. All auto-indexed with 92% confidence.
11:30 AM · Slack Thread
DocBrain answers with fragments captured 2 hours ago, plus links to the original PR. The support engineer types /docbrain capture to save more context.
2:00 PM · Auto-Composition
DocBrain detects semantic similarity across PR fragments, Slack capture, and older fragments. Composes "Payment Refund Procedures," scores at 78/100, routes for SME review.
3:30 PM · Published
Quality rules catch 2 violations, auto-fixed. Final score: 91/100. Published to "Payments" space. Total engineer effort: zero.
Traditional: schedule a doc sprint, assign writers, review in 2 weeks, stale in 2 months.
Shift-left: captured at 9:15 AM. Published by 3:30 PM. Stays accurate forever.
Independent Assessment
No cherry-picked benchmarks. We opened Rovo, asked it to do a fair comparison, and published the full response — including where it wins.
DocBrain's 'capture knowledge that was never written down' is solving a problem I fundamentally can't.
— Rovo (Atlassian's AI assistant), responding to a direct comparison prompt
Full response published unedited. Rovo's assessment, not ours. Read the full transcript →
FAQ
bash scripts/setup.sh and it brings up PostgreSQL, OpenSearch, Redis, migrations, and sample docs. Add your LLM API key and go.Security & Access
Docker, an API key, and three commands. That's it.