Skip to content

Lantern

Stop digging through Confluence and Jira. Ask a question, get the page you meant.

Lantern is a self-hosted retrieval engine over your company's Confluence and Jira. It reads every page, spec, and ticket, understands what they mean, and hands the right ones to you — or to your AI coding agent — in one call. No more keyword roulette, no more "I know we wrote this down somewhere."

It plugs straight into Claude, Cursor, Codex, and Claude Code over MCP, so the agent you already code with can pull internal knowledge mid-task — and even write back to Confluence and Jira under your name.

See it in action

You ask in plain language. Lantern ranks by meaning, not keywords, and returns the page with its breadcrumb trail and a deep link.

Natural-language search — you don't know the title, just the idea:

"how do we retry failed payment webhooks?" → Payments › Webhooks › Retry & idempotency — the design section, top hit, with a link straight to the heading.

Jira issue, instantly — paste a key, skip the search:

"what's the status of CMP-1422?" → the issue, its current state, assignee, and the linked epic — fetched directly, no ranking needed.

Cross-source, one question — the spec lives in Confluence, the work in Jira, the service in Backstage. Lantern follows the links between them:

"everything about the Nova Poshta delivery integration" → the PRD page + its epic and open tickets + the backreferences from other teams that depend on it — gathered in a single context pack.

Multilingual out of the box — ask in English, find the Ukrainian doc (and the other way round). The embeddings are language-agnostic, so the team's mix of EN/UK/RU just works.

From inside your coding agent

Because Lantern speaks MCP, the knowledge shows up where you're already working:

You (in Claude Code): "Before I touch the checkout flow — how is our payment-method model supposed to work? Pull the spec."

Claude: calls Lantern search → context_pack … "Per Payments › Data model for all payment methods, every method maps to a single PaymentMethod entity with… and the open epic CMP-1390 is reworking the display layer — here's the doc and the three linked tickets."

The agent reads the real internal design before it writes code — instead of guessing or asking you to paste links.

And with your Atlassian PAT connected, it goes both ways: the agent can update a section, comment on an issue, or transition a ticket — every write attributed to you, gated by your permissions, and version-checked so it never stomps a teammate's edit. Dry-run first, always.

Why it finds what grep can't

Lantern is a real retrieval stack, not a search box bolted onto a wiki:

  • Semantic, not keyword — multilingual embeddings rank by meaning, so it finds the page about "idempotency keys" when you asked about "duplicate charges", and matches across EN / UK / RU without you translating anything.
  • Exact when it counts — paste a Jira key and it short-circuits straight to the issue; ambiguous titles get disambiguated instead of buried.
  • Graph-aware — it knows which pages link to which, which tickets implement which spec, and which services a doc describes. Ask about one thing, get its whole neighborhood (parents, mentions, backreferences, shared repos / projects / spaces) — and, when you pull the full context pack, two-hop jumps through what they share.
  • Recency-tuned — fresher docs surface above stale ones, so you get this quarter's decision, not the 2-year-old draft.
  • Permission-scoped — you only ever see what your own Atlassian account is allowed to read. Lantern never leaks a doc across team boundaries.

Public vs. signed-in

This wiki is public — you are reading it right now without an account. Everything under /wiki/ (this documentation) is served without authentication.

The Lantern app is not. Search, the doc graph, context packs, and every write tool live behind sign-in: you authenticate with your email and password (or your Atlassian PAT), the browser gets an HTTP-only session cookie, and AI tools authenticate with a token. The split is enforced server-side — anonymous requests to anything but the docs are rejected.

Surface Account needed?
This wiki (/wiki/…) No
Search / browse / context pack (web UI + API) Yes — sign in
Personal access tokens, API keys Yes — sign in
AI tool (MCP) access Yes — token

I have access to Lantern at work — what now?

  1. Sign in & see your docs — open /login and either paste your Atlassian PAT ("Sign in with Atlassian" — no Lantern password) or use your email + password. Connecting your Atlassian PAT is what makes search show the docs you can read — the guide walks you through it.
  2. Connect your AI tool (this is where it gets good) — one download wires Lantern into Claude Desktop, Cursor, Codex, or Claude Code so your agent can query and edit internal knowledge while you work.
  3. Mint a long-lived token (advanced) — for scripts or token-based AI configs, from the Tokens page.

Lantern is deployed inside your company network. You do not need to install or run a server — your team admin already did that. You only configure the client side.

Need admin access or to deploy Lantern for your team?

This wiki is for users of Lantern. If you are an operator setting up an instance, see the internal repository documentation (docs/internal/ in the source tree).