camunda-docs

Camunda Docs Lookup

Two retrieval paths into docs.camunda.io, each with a different strength. Pick by question shape; fall back to llms.txt only if both are unavailable.

Path	Mechanism	Best for	Returns
camunda-docs MCP	Semantic retrieval (vector search) over docs + other Camunda knowledge sources	Conceptual or fuzzy questions; you don't know the exact term; want chunks ready to read	Top-k markdown chunks with source URLs
Algolia DocSearch	Classical lexical search (the same index docs.camunda.io's own search box uses)	You know the term; want version-scoped results (8.7 vs 8.9 vs next); want to triage a ranked page list	Ranked pages with title, hierarchy, snippet — fetch the page for full content
llms.txt / llms-full.txt	Static text index/corpus	Offline fallback when neither service is reachable	Plain text; grep locally

Use both primary paths in parallel when in doubt — they're cheap and frequently surface different relevant pages.

camunda-docs MCP server

If the camunda-docs MCP server (https://camunda-docs.mcp.kapa.ai, HTTP transport) is connected, call its knowledge search tool with a single, well-formed natural-language sentence (the tool requires a complete sentence, not keywords).

If it isn't connected, suggest installing it once so the user has it set up for future questions. Skip the suggestion if you've already made it earlier in this conversation. For Claude Code:

claude mcp add --transport http camunda-docs https://camunda-docs.mcp.kapa.ai

For VS Code Copilot, Cursor, or generic MCP clients, see the reference. Google sign-in on first use; 40 requests/hour and 200/day per user; not for CI.

Algolia DocSearch

A bash wrapper around the public Algolia DocSearch endpoint lives at scripts/docs-search.sh. Run it as:

scripts/docs-search.sh "<query>" [--version stable|next|<major>.<minor>|all] [--limit N]

Default --version stable resolves to the highest numeric version in the index (so it keeps working when 9.0 ships). Default --limit 10. Returns JSON:

{
  "query": "...",
  "version": "stable",
  "algolia_version": "8.9",
  "total": 5,
  "total_matching": 25,
  "hits": [
    { "url": "...", "hierarchy": "...", "title": "...", "snippet": "..." }
  ]
}

• total — hits in this response.
• total_matching — raw count of pages matching the query in the index, before pagination/dedupe. Useful for "how broadly does my term appear?"
• algolia_version — the version facet value sent to the index (e.g. "8.9", "current"). null in --version all mode (no facet filter applied).

Search flow

1. Pick a version (see table below). Default stable.
2. Run the script.
3. Inspect the hits — each has url, hierarchy, title, snippet.
4. Fetch the most relevant page(s) directly by URL to read full context. Snippets are short; don't answer from them alone if the question needs detail.
5. Refine if insufficient — see refinement strategy below.

Version handling

The Camunda docs are versioned in the URL path. The Algolia version facet exposes those versions — without filtering, the same page returns 3-4 times across versions, which is noisy.

Flag value	Algolia facet	URL prefix	Use when
stable (default)	highest numeric (e.g. 8.9)	/docs/	General questions, no version context
next	current	/docs/next/	Upcoming/unreleased behavior, alpha features
8.8 / 8.7 / etc.	matching numeric	/docs/8.8/ etc.	Customer/cluster on a specific generation
all	(none)	any	Comparing versions; results deduped by URL stem so best-relevance version per page wins

Pick explicitly when:

• Question is about a specific customer/cluster on a known generation → use that version.
• Question is about an upcoming feature, an alpha, or "will this change in 8.10?" → run stable and next in parallel and compare.
• Question is version-neutral (concepts, BPMN semantics, generic API shape) → stable is fine.

In --version all mode the script over-fetches (4× --limit, capped at 100) so that after dedupe by URL stem the response contains close to --limit unique pages. Dedupe preserves Algolia's relevance order — first hit per unique URL stem wins, then trimmed to --limit.

Examples

# Default — latest stable
scripts/docs-search.sh "zeebe gateway long polling timeout"

# Specific version (customer on 8.7)
scripts/docs-search.sh "task assignment" --version 8.7

# Upcoming behavior
scripts/docs-search.sh "agentic connectors" --version next

# Compare across versions (deduped by URL stem)
scripts/docs-search.sh "decision evaluation API" --version all --limit 15

Refinement strategy

If the first query is unhelpful:

• Too broad → too many off-topic hits: add specifics (component, error code, config key).
• Too narrow → no hits: drop adjectives, try the canonical product term (e.g. "Tasklist" not "task UI", "Zeebe" not "workflow engine").
• Wrong version: try --version all to see if the term exists elsewhere, then narrow to the right release.
• Synonyms: try alternative names — e.g. "Operate" ↔ "process monitoring", "Optimize" ↔ "analytics".

Don't loop more than 2-3 refinements. If still nothing, switch to the MCP path (or note the gap to the user).

No credentials needed

The Algolia search-only API key is public — Algolia ships it in the browser JS bundle of docs.camunda.io, so every visitor already has it. It's hardcoded at the top of scripts/docs-search.sh with a comment explaining why it's safe to commit. No env vars, no auth, safe in CI.

Without the wrapper script

If you can't or don't want to run scripts/docs-search.sh (e.g. jq missing, restricted sandbox, or you just need the raw API), call the Algolia endpoint directly:

curl -sS -f -X POST \
  "https://6KYF3VMCXZ-dsn.algolia.net/1/indexes/camunda-v2/query" \
  -H "X-Algolia-API-Key: 68db7725a8410eace68419c29385ad1e" \
  -H "X-Algolia-Application-Id: 6KYF3VMCXZ" \
  -H "Content-Type: application/json" \
  -d '{"query":"YOUR QUERY","hitsPerPage":10,"facetFilters":[["version:8.9"]]}'

Substitute the version: facet for the version you want (current for next, omit facetFilters entirely for all). The response shape is Algolia's raw hits — useful fields per hit are url, hierarchy.lvl0..lvl5, and content. If curl itself isn't available either, use any HTTP client that can POST JSON (Python urllib, Node fetch, PowerShell Invoke-RestMethod); the endpoint, headers, and body are the same. The credentials in the snippet are public by design (see "No credentials needed" above).

Fallback: llms.txt (docs index)

If neither MCP nor Algolia is usable, fetch https://docs.camunda.io/llms.txt (~540 KB) — an index of - [Title](url): description. lines for every Camunda 8 docs page. Do not read the whole file into context. Cache it locally, search for the term to find the matching URL, then fetch that page directly.

Last resort: llms-full.txt (full docs corpus)

Only if llms.txt doesn't surface what you need, fall back to https://docs.camunda.io/llms-full.txt (~11 MB) — every docs page concatenated. Even larger; never read whole into context. Cache locally, search with surrounding context for the term. Use when the right page name isn't obvious from the index.