Langdock Integration Research

Date: 2026-05-11 Researcher: Claude (api-integration-research skill) Source URL: https://langdock.com/de/products/api

TL;DR

Langdock is a Berlin-based, GDPR-compliant AI gateway that proxies major LLM providers (OpenAI, Anthropic, Google, Mistral) through an EU-hosted layer plus its own product surface (chat, agents, knowledge folders, workflows). For Orcha there are three viable integration angles:

1. Use Langdock as Orcha's LLM provider for EU customers — drop-in via OpenAI/Anthropic SDK base URL swap, 10% markup, EU residency built-in.
2. Expose Orcha as a Langdock-callable tool — Langdock supports MCP servers and third-party integrations with 754 native actions; an Orcha MCP server would let Langdock agents push/pull invoice data.
3. Trigger Orcha actions from Langdock workflows via webhooks — Langdock workflows have webhook nodes, so customer-built workflows could call Orcha endpoints.

Angle 1 is the lowest-effort, customer-visible play. Angle 2 is the strategic play (Orcha-as-a-tool inside a customer's existing AI portal).

1. Summary — Capability Matrix

2. API Landscape

Base URLs

• Default: https://api.langdock.com
• Dedicated enterprise deployment: <deployment-url>/api/public
• Regions in path: /openai/eu/, /openai/us/, /anthropic/eu/, etc. — region is part of the path, not the host.

Endpoint groups (from docs.langdock.com/llms.txt)

• Completions — OpenAI, Anthropic, Google, Mistral (each as a provider-compatible endpoint)
• Embeddings — OpenAI-compatible
• Agents — full CRUD: create/get/update/disable/publish/migrate; plus attachment upload
• Assistants — legacy, migration path to Agents documented
• Knowledge Folders — upload, retrieve, search, update, delete attachments; sharing
• Integrations — create/list/get/update/delete integrations, actions, triggers; update auth
• Audit Logs — list
• Usage Export — agents, models, projects, users
• User Management — invite, activate, deactivate

Authentication

• Bearer token (Authorization: Bearer <KEY>) or x-api-key header (MCP)
• Keys are workspace-scoped, with per-key scopes (Completion API, Agent API, etc.)
• Created in workspace settings; browser-origin requests blocked by default

Rate Limits

• Default: 500 RPM, 60,000 TPM — applied at workspace level per model
• Higher limits on Enterprise on request

SDK Strategy

No first-party SDK. Langdock explicitly recommends using the vendor's official SDK with a custom base_url:

# Anthropic SDK pointing at Langdock EU
client = Anthropic(
    base_url="https://api.langdock.com/anthropic/eu/",
    api_key="<LANGDOCK_API_KEY>"
)

// OpenAI SDK pointing at Langdock EU
const provider = createOpenAI({
  baseURL: "https://api.langdock.com/openai/eu/v1",
  apiKey: process.env.LANGDOCK_API_KEY,
});

Vercel AI SDK is mentioned as natively compatible for the Agents API.

Unsupported OpenAI features (worth noting for Orcha)

The OpenAI-compatible endpoint accepts everything except: n, service_tier, parallel_tool_calls, stream_options. Most are not relevant for Orcha's extraction workloads.

Prompt caching

Not documented on the Anthropic-compatible endpoint. Orcha doesn't use caching today (per MEMORY.md), but if we adopt it on the direct Anthropic API we should retest whether Langdock passes cache_control blocks through to upstream. Confidence: low — needs empirical verification.

3. Write Verification

Write capability confirmed across all relevant data objects:

These are real writes (creating records), not POST-shaped reads. Confidence: high.

4. iPaaS Findings

Make.com — native connector exists ✅

4 actions, no triggers:

1. Create a Codestral (write — Mistral model invocation)
2. Create an OpenAI Chat Completion (write)
3. Send an Anthropic Message (write)
4. Make an API Call (generic — read/write any Langdock endpoint)

Use case for Orcha: customers can wire Langdock into Make scenarios alongside Orcha-touched tools. The generic "Make an API Call" module means any of the agent/knowledge/integration endpoints are reachable from Make even without a dedicated module.

Zapier — no native connector found.

n8n — no native connector. Community guides describe configuring n8n's OpenAI node with Langdock's baseURL as a compatibility hack. Not officially supported; minor features (e.g. unsupported OpenAI params) break.

Workato / Celigo / Tray — none surfaced.

Native Langdock integrations directory (not iPaaS, but relevant)

Langdock ships 754+ native actions across: Slack, Teams, Gmail, Outlook, Google Drive/Docs/Sheets, OneDrive, SharePoint, Notion, Confluence, Jira, Asana, Linear, Monday, Salesforce, HubSpot, Stripe, Zendesk, BigQuery, Snowflake, Databricks, Metabase, Looker, Power BI, Tableau, GitHub, ElevenLabs, DeepL, ServiceNow, Personio, Ashby, Airtable, Calendly, Luma, Pinecone, Qdrant, Milvus, Vertex AI, Azure AI Search, AWS Kendra.

Orcha is not listed. Building a native Langdock integration (or MCP server) would be net-new.

5. Alternative Channels

6. Licensing & Access Requirements

What the customer must buy/enable

Sample API rates (10% markup confirmed)

• Claude Opus 4.6: €4.70 input / €23.49 output per 1M tokens
• GPT-5.2: €1.64 input / €13.15 output per 1M tokens
• Compare to direct Anthropic Opus 4.6: ~$4.27 input / $21.35 output → roughly +10% in EUR.

Ongoing costs for the customer

• Per-seat workspace fees (only if using chat/agents product)
• Per-token API consumption with 10% markup vs. direct provider
• Workflow run consumption (if using workflows)
• No transaction packages, no per-record charges

7. Orcha-Specific Deep Dive

Angle A — Langdock as Orcha's LLM provider (EU customer story)

The proposition: For German/Swiss customers who want EU data residency on LLM calls, point Orcha's workers/llm.clj at Langdock instead of api.anthropic.com.

What changes in Orcha:

• base_url → https://api.langdock.com/anthropic/eu/
• API key → customer's Langdock key (BYOL) or Orcha's shared key
• Anthropic SDK call shape stays the same
• All current params (messages, system, max_tokens, tools, tool_choice, thinking, temperature) supported

Unverified gaps:

• Prompt caching pass-through — not documented; need to test cache_control blocks if Orcha turns caching on
• Latency — extra hop EU-customer → Langdock EU → Anthropic (likely AWS Bedrock EU under the hood). Not benchmarked.
• Per-customer billing — if Orcha uses one shared Langdock key, Orcha eats the 10% markup; if customer provides their key (BYOL pattern), Orcha just routes
• Cost reporting — Langdock CSV usage export gives per-API-key, per-model token usage, format matches Orcha's existing acquisition_llm_stat shape

Confidence: high that the technical swap works. Medium on caching and latency until tested.

Angle B — Orcha as a Langdock-callable tool (strategic distribution)

The proposition: Customer uses Langdock as their company AI portal. We expose Orcha as either an MCP server or a Langdock-native action so an employee chatting in Langdock can say "pull all unpaid invoices from supplier X" and Langdock's agent calls Orcha.

Two implementation paths:

1. Orcha as MCP server — easier. Langdock auto-discovers MCP tools and converts them to actions. Auth via OAuth 2.0 (DCR or non-DCR) or API key. Lower distribution barrier; works for any MCP-compatible client (Claude Desktop, Cursor, ChatGPT) — not Langdock-exclusive.
2. Native Langdock integration — harder. Langdock's integration directory uses create-integration API to register custom actions/triggers, but the integration must be listed via Langdock partnership for general availability. Better visibility, more lock-in.

Capabilities to expose:

• find_supplier_invoices (read)
• get_invoice_status (read)
• approve_invoice (write — needs careful scoping)
• push_invoice_to_erp (write)
• search_documents (read — leverages Orcha's semantic search)

Confidence: high that MCP path is feasible. Effort: build + host an MCP server fronting Orcha's existing API + handle OAuth.

Angle C — Trigger Orcha from Langdock workflows

The proposition: Customer builds a workflow in Langdock that, e.g., processes inbound supplier emails, drafts a categorization, and pushes the result to Orcha. The workflow ends in an HTTP Request node calling an Orcha endpoint.

What works today: Langdock workflows have an HTTP Request node — they can call any Orcha endpoint without us building anything. Conversely, Orcha could call a Langdock workflow's webhook trigger.

Confidence: high — no integration work needed, but also lowest strategic value (customer-side glue, not an Orcha product capability).

Orcha Integration Capability Summary

8. Recommended Path Forward

If the goal is EU GDPR story for Orcha customers (most likely):

Phase 1 — proof of concept (1–2 days)

• Spike: route a single Orcha extraction prompt through Langdock's Anthropic endpoint
• Measure latency delta vs. direct Anthropic
• Verify tool-use round-trip works for our existing extraction tools
• Test if cache_control blocks pass through (when we eventually adopt caching)

Phase 2 — config switch (1 week)

• Add :provider :anthropic-via-langdock option in workers/llm.clj with separate base URL + key
• Per-tenant config in admin so customers opting in get routed via Langdock
• Cost reporting: import Langdock's CSV export into acquisition_llm_stat (or model after it)

Phase 3 — sales packaging

• "EU data residency on every LLM call" as a feature for German/Swiss prospects
• BYOL option (customer brings their Langdock key) so we don't eat the 10%

If the goal is strategic distribution (longer-term play):

Build an Orcha MCP server that exposes the customer-relevant invoice/document operations. Test it first via Claude Desktop, then submit to Langdock's MCP directory. This makes Orcha usable from inside any AI portal — Langdock, Claude, Cursor, ChatGPT — not just one.

What I would NOT do:

• Build a native Langdock integration before MCP is proven — MCP is the wider standard, native Langdock is single-vendor lock-in.
• Adopt Langdock as Orcha's primary LLM gateway for all customers — the 10% markup compounds and we lose direct provider relationship leverage. Use it as an EU-residency option, not the default.

9. Sources

Official Langdock

• API product page
• Pricing
• Models
• Integrations directory
• Workflows
• Changelog
• Trust portal

Documentation

• Docs root
• Full docs index (llms.txt)
• OpenAI completion endpoint
• Anthropic completion endpoint
• Agent API guide
• Knowledge folder upload
• Knowledge folder search
• MCP integration (Langdock as client)
• MCP server (Langdock as server)
• Webhook trigger node
• API usage export
• Pricing details

Community & iPaaS

• Make.com Langdock connector
• Langdock GitHub org
• n8n compatibility tutorial (decaid.ai)

Independent reviews

• tldv.io review
• Cybernews review
• Workativ pricing breakdown