Claude for Work vs Langdock

Claude for Work

The commercial terms state plainly that Anthropic may not train models on Customer Content, Enterprise admins can set retention for conversations, projects and artifacts with every change tracked in audit logs, and the abuse-monitoring carve-outs are disclosed rather than buried. I hold it below the top because data is retained indefinitely unless an owner acts, and the retention controls explicitly do not reach Claude Managed Agents or cloud-run Claude Code — the agent features I would deploy first.

Langdock

The terms commit that neither Langdock nor any model provider it engages may train on customer content, Azure models run with modified abuse monitoring described as zero data retention for prompts and outputs, and a DPA under Art. 28 GDPR is automatically concluded. Admins set the chat retention period, deletion behaviour is documented per feature for chats, agents, folders and knowledge bases, and data is deleted within 30 days of termination. I would want the same abuse-monitoring and zero-retention detail for every provider, not just Microsoft, before calling it fully accounted.

Claude for Work

The model families are named — Mythos including Fable, Opus, Sonnet, Haiku — with per-role control over which models members may use, and the privacy policy states personal data is transferred to servers in the US or other countries outside the EEA and the UK. We found no public information on pinning this product's inference or storage to the EU; the regions page captured simply hands off to Amazon Bedrock's own documentation.

Langdock

EU-only processing is the default and inference leaves the EU only if a customer actively selects a global-deployment model, with data at rest still held in EU data centres and Frankfurt hosting under the Azure EU Data Boundary documented. Forty-plus models are offered across named providers, bring-your-own-key locks a workspace to only customer-configured models, and single-tenant, own-cloud and on-prem (Kubernetes via Helm) deployments are documented from 2,000 to 5,000 seats. A published per-model inference-region table and open-weight or self-hosted model support are not documented on the captured pages.

Claude for Work

This is what I want before wiring an agent into company systems: SSO with SAML and domain capture, SCIM provisioning, custom roles with per-connector permission levels — every tool 'Always allow', 'Needs approval', 'Blocked' or per-tool 'Custom' — plus model access per role and per-group monthly spend limits. The gaps sit at the regulated end: we found no public information on data-loss prevention rules, streaming audit logs to a SIEM, or per-agent ownership, and the scope of the usage analytics is not described.

Langdock

SAML 2.0 SSO with any compatible identity provider, SCIM, signed authn requests, domain verification, enforced MFA at the identity provider and IP restrictions give IT real control, and retrieval mirrors source-tool permissions one to one so an agent cannot read beyond its user. Usage analytics can be enabled or disabled per workspace with CSV and API exports of cost and token data. The governance module lets admins review, approve, flag or disable agents with compliance rules re-checked on every published version; we found no public information on exportable audit logs or streaming to a SIEM.

Claude for Work

Remote MCP connectors, a documented API, webhooks and a marketplace with an official Pipedrive connector carrying seventy-nine read-and-write tools that respect existing CRM permissions — that is real plumbing, not slideware. The library as documented leans on mail, drives and collaboration tools; we found no public information on ticketing connectors, sync frequency or completeness, or how grounded answers cite their sources.

Langdock

MCP is first-class: a directory of 79 officially operated remote servers, custom MCP servers connectable with OAuth 2.0 or API-key auth, MCP tools surfaced as actions with confirmation flows, MCP Apps rendering interactive UI in a sandboxed iframe in chat, an agent-to-agent protocol, plus a unified API, native connectors, Office add-ins, desktop and mobile apps and Zapier reach into 7,000 apps. We found no public information on incremental sync behaviour, webhooks, or source citations on grounded answers, which I need before pointing this at ERP and CRM flows.

Claude for Work

Agents run multi-step sessions that act on your behalf in connected systems, run events stream on the session so you can follow each step, and the documentation is unusually honest about limits — a thousand agents per run, sixty-four concurrent threads, a twenty-four-hour run lifetime, a failed agent usually retried once on a new thread. What stops me signing off: the docs state there is no approval step before a managed-agent run starts, though connector tools can be gated so members confirm each call, and we found no public information on test or evaluation tooling to rehearse an agent before the whole company uses it.

Langdock

This is what I look for: action nodes with human confirmation so nothing is sent without approval, workflow pauses with approve, reject and form input, conditions so only some runs need a person, a workflow monitor for pending steps, and a builder test toggle with test runs that skip notifications. Agents get versioned publishing where later edits stay drafts until published, scheduling, scoped write folders, subagents, Slack and Teams channels, and a governance view exposing each agent's instructions, integrations, costs and history with the power to disable it until fixed. Evaluation against defined test sets and documented failure handling beyond a general output disclaimer are not documented.

Claude for Work

EEA, Swiss and UK customers contract with Anthropic Ireland Limited, with Dublin arbitration and standard contractual clauses covering transfers, but the privacy policy states personal data is processed on US servers, so hosting is not European. A subprocessor list is published as a reference, though the pages captured name only two Korean processors, and we found no public information on EU inference or an EU hosting option for this product.

Langdock

Langdock GmbH in Berlin (registered under HRB 254499) is the contracting entity under German law with exclusive jurisdiction in Berlin, hosting runs on Microsoft Azure in Frankfurt with the EU Data Boundary, and a published subprocessor list names every model provider with its location, including OpenAI Ireland, AWS, Google Cloud and Azure, plus Sentry in the US covered by SCCs and Data Privacy Framework safeguards. Transfers outside the EU happen only on the customer's instruction, and ISO 27001:2022 and SOC 2 Type II certificates are published. We found no public information on the company's ownership structure, so I credit the dimensions the evidence proves.

Claude for Work

The headline numbers are public: the free plan at $0, Pro at $17 per month with annual subscription discount ($200 billed up front) or $20 if billed monthly, Max 'From $100 Per month', and Enterprise at $20 per seat per month plus usage billed at API rates with a minimum of 20 seats and an annual commitment. A buyer still cannot total a year from these pages alone: usage limits are flagged but not quantified, the per-token rates live on a separate pricing page, and we found no public information on a Team plan price, overage rates or add-ons.

Langdock

Public per-seat prices of €23.20 standard and €99.00 Business Max with 5x usage (both excluding VAT, 20% off annual), workflow tiers with 2,500 runs included and €449 per workspace per month for 40,000 runs, a €2.80 governance add-on free until January 1, 2027, and per-token API rates make most of an annual bill computable. The included usage quantity per seat rests on a fair-use policy rather than published numbers, overage is not priced, the exact governance add-on price varies by seat count without a published scale, and Enterprise pricing is on request.