Langbase

Pipe and Agent primitives support function calling with custom tools and structured outputs, the Tools primitive adds web search and crawling, and agents connect to third-party MCP servers such as Slack or Intercom through an mcp_servers parameter, with Composio and n8n integrations documented.

SourceLangbase, langbase.com/docs mcp-servers, tools and features/tool-callingread 2026-09-21

The Workflow primitive runs multi-step agentic applications as defined steps, sequential or parallel, with conditional execution paths based on earlier step results, configurable retry strategies, timeouts and step-by-step logging of the execution flow. The buyer configures the branching, retries and routing across those steps.

SourceLangbase, langbase.com/docs/workflowread 2026-09-21

Memory is a managed retrieval layer: uploaded documents are queued, processed, chunked and embedded into vector storage, a Pipe retrieves query-relevant passages by similarity with reranking through a customizable RAG prompt, and documents can be replaced or re-embedded as they change. That is an index over the customer's knowledge that persists and stays queryable.

SourceLangbase, langbase.com/docs memory/concepts and explained/ragread 2026-09-21

Pipes call a moderation endpoint to identify harmful content for OpenAI models, on by default, and a safety prompt defines instructions such as not answering outside the given context; these are structured guardrails. No approval step, escalation rule or pause for a person's sign-off before an agent acts is documented.

SourceLangbase, langbase.com/docs features/moderation and features/safetyread 2026-09-21

Access controls are published: organizations with member access and organization-level keys, keysets that scope provider keys per user, organization or pipe, and SAML single sign-on and RBAC access controls on the custom plan. The security page states SOC 2 Type II compliance with the latest audit currently under review, which names no report, auditor or date, so the compliance claim is asserted rather than attested.

SourceLangbase, langbase.com/security, pricing and docs features/organizations and features/keysetsread 2026-09-21

Every Pipe request produces a log in the Usage tab with the pipe version, type, model, prompt and completion tokens, cost and cache status, workflows log their execution step by step, and log retention is one week on Individual and Growth and unlimited on the custom plan. The documented log fields are run metadata; no way to inspect prompts, tool calls, retrieved knowledge and outputs step by step is documented.

SourceLangbase, langbase.com/docs features/logs and workflow, and pricingread 2026-09-21

Threads store conversation messages and metadata through an API that creates, appends to, lists and deletes threads, so a conversation's history and context persist across requests; Pipes can also store messages. The documentation states no lifetime for thread data and no memory scoped to a user or team beyond the thread.

SourceLangbase, langbase.com/docs threads and features/store-messagesread 2026-09-21

Langbase runs pipes, memory and workflows as a managed serverless cloud, and an agent built with Command can be downloaded for the customer to self-host, a one-sentence self-hosting route. No self-hosted Langbase runtime, private cloud or region selection is documented.

SourceLangbase, langbase.com/docs/command and pricingread 2026-09-21

A customer can fork ready-made Pipe agents from Langbase into their own account and run them, among them an AI chatbot, a tech guide writer, a bug report summarizer, a software docs translator, a resume analyst and an API security consultant, and Command lets users fork other agents and make them their own. Each is a whole agent doing its own job.

SourceLangbase, langbase.com/docs pipe/examples, features/fork and commandread 2026-09-21

Pipes, agents and workflows run when the developer's application calls their API, and guides such as the Discord bot and email agent show the developer's own code doing the invoking; Langbase's documentation describes no schedule, webhook or event source that wakes an agent without a person or the developer's code initiating it.

SourceLangbase, langbase.com/docs (full documentation index)read 2026-09-21

The customer chooses the model for each Pipe or Agent from more than 100 models across OpenAI, Anthropic, Google, Together, Fireworks, Groq, DeepSeek, Perplexity, Mistral, Cohere, xAI, Azure OpenAI and OpenRouter, using its own provider keys through keysets set per user, organization or pipe, with model presets.

SourceLangbase, langbase.com/docs supported-models-and-providers and features/keysetsread 2026-09-21

Langbase documents a REST API reference and an SDK for pipes, memory, threads, tools and workflows, authenticated with user or organization API keys, and runs a remote MCP server at mcp.langbase.com through which outside assistants work with a customer's pipes and memories.

SourceLangbase, langbase.com/docs api-reference, sdk and mcp-servers/remote-mcp-serverread 2026-09-21

Experiments rerun the latest Pipe configuration against the last five generate requests and show previous and new completions side by side, so a builder can inspect how a change affects responses, and pipes are versioned; evals are listed on the custom plan without further documentation. No scored test cases, evaluation harness or quality gate returning a verdict is documented.

SourceLangbase, langbase.com/docs features/experiments and features/versions, and pricingread 2026-09-21

No browser, desktop or computer control is documented. The Tools primitive's crawling and search run on third-party services the customer holds its own account and key with, and a third-party browser rendering MCP server is mentioned only as an example of what can be connected.

SourceLangbase, langbase.com/docs tools and mcp-serversread 2026-09-21