Albato

One MCP endpoint reaches more than 1,000 apps and about 20,000 actions, and every call runs under the end user's own connected account. The agent finds an app, its actions and the user's connections in a single batched call, fetches the action's input fields and runs it synchronously, while Albato stores and refreshes OAuth tokens and routes each request to the right tenant. When an account is missing, the agent sends a white label connect link. Custom HTTP requests also work as tools, and Albato builds two custom connectors a month on Pro.

Sourcealbato.com/wiki/articles/mcp-tools-referenceread 2026-11-09

Agents run as steps inside Albato automations, where a workflow engine with retries, idempotency and schedules sequences actions and a router splits a flow into if and else branches. An AI Agent step reads data from earlier steps, decides which of its tools to call, and hands its output to the fixed steps that follow. Copilot builds, validates and starts whole automations on request, so a partner's agent can delegate a multistep flow rather than chain each call itself. Paid plans run scenarios in parallel.

Sourcealbato.com/wiki/articles/ai-agent-how-to-set-it-upread 2026-11-09

Grounding comes only from the data an automation passes in from earlier steps and from whatever the agent's tools return at run time. The AI Agent step cannot connect a knowledge base, and grounded decisions over a knowledge base or trusted external sources are still marked coming soon for embedded agents. Copilot answers questions about Albato from the vendor's own help content, which covers the product rather than the customer's material. Albato keeps no index, vector store or document library over a customer's files.

Sourcealbato.com/wiki/articles/ai-agent-how-to-set-it-upread 2026-11-09

Each agent gets only the tools and actions a builder grants it, plus a guardrails field of up to 1,000 characters stating what it must not do. Field locks keep critical values such as email addresses, billing amounts or deal stages out of the agent's reach, and partners can limit tools per tenant. Copilot stops to ask the user which action or account to use and leaves a finished automation for review before launch. Routing high impact actions to a person for approval in Slack or email is marked coming soon for embedded agents.

Sourcealbato.com/embedded/ai-agentsread 2026-11-09

Embedded Enterprise adds workspaces with role based access control and granular permissions over who can create, edit, launch or manage integrations, and the self serve Teams plan carries roles, invites and spaces. Each agent gets its own revocable virtual key with an optional expiry, the real session token stays encrypted on the server, and each end user's OAuth tokens sit apart from every other tenant's. Albato holds a SOC 2 Type 2 attestation with annual audits and the report under NDA, runs third party penetration tests and is working toward ISO 27001.

Sourcealbato.com/embedded/pricingread 2026-11-09

Every MCP call, retry and error lands in one dashboard across all apps and users, and each agent action is logged with 60 day retention and a GDPR opt out. The automation log shows what data came in, what the agent decided and what it did, so a team can trace a wrong action back to its instructions. The partner portal reports usage by user, app, trigger and action with a live health rate, and alerts for integration errors and transaction limits go out by email or Slack.

Sourcealbato.comread 2026-11-09

Memory of previous runs is off by default on the AI Agent step and can be switched on per agent. Memory size sets how many recent interactions it considers, from 1 to 100, where one input and one output make one interaction, and a thread ID keeps each user or conversation separate, such as one Telegram chat per customer. No retention period or purge control comes with that history. Copilot keeps context within a conversation ID, and the MCP server is task based, so an agent holds no state there between calls.

Sourcealbato.com/wiki/articles/ai-agent-how-to-set-it-upread 2026-11-09

Data lives on AWS in the US by default, across two availability zones, and Enterprise customers can keep it in the EU, the UK or Canada instead. Albato also runs a managed private AWS cloud per customer, included on Enterprise and an extra fee below it, and self hosting places the embedded platform in the customer's own infrastructure or on premises with continued updates. The MCP server can be hosted by Albato or by the customer, and embedded agents can run on premises against local models.

Sourcealbato.com/embedded/pricingread 2026-11-09

Ready made agents come as starting points in the self serve product, among them an Ads Performance Analyst that watches cost per click and cost per lead for anomalies, a Content and SMM Agent, an SEO Agent and a Reporting Agent that writes campaign results to Google Sheets on a schedule, each with its own job. Embedded partners package agents as reusable templates for common customer workflows, and Solutions bundle frequent use cases into one click integration kits, unlimited on every Embedded plan. Agencies clone agent automations across client accounts.

Sourcealbato.com/ai-agents-marketingread 2026-11-09

Work reaches an agent through the trigger that starts its automation: a webhook, a schedule or an event in any connected app, such as a new HubSpot contact or a Zendesk ticket. Albato listens for triggers and polls APIs as often as every minute on Embedded Pro, without charging for either, and sets up custom webhooks for apps that lack an API. Partners subscribe their users to the partner product's own webhooks through the Embedded API, so an event inside that product can start an agent run.

Sourcealbato.com/ai-agentsread 2026-11-09

Builders pick the provider for each AI Agent step from OpenAI, Anthropic Claude, Google Gemini, DeepSeek or the built in Albato AI, then enter the model ID, and can switch at any time. Embedded agents take a model per workflow, including a partner's own model, Bedrock, Azure or a self hosted model, so cost, quality and compliance can differ from step to step. Customers can move from ChatGPT to Claude or Gemini in one click, and runs on their own provider key carry no token charge from Albato.

Sourcealbato.com/wiki/articles/ai-agent-how-to-set-it-upread 2026-11-09

Partners drive the platform through the Embedded API, 11 REST methods that add, block and delete users, refresh their tokens, set and list billing plans, top up transactions, and create or read connections to the partner's own system. Server calls use a permanent API key, and each end user gets an expiring JWT session token for the embedded builder. Copilot is callable over A2A version 1.0 at a JSON RPC endpoint, with a public agent card, a partner token paired with the user's session token, streamed task events, and methods to send, check and cancel tasks.

Sourcealbato.com/embedded/apiread 2026-11-09

Pro and Enterprise include a staging environment, isolated from production and syncable to the partner's sandbox, for testing integration solutions before release. Builders tune an agent by adjusting its instructions and tightening its guardrails after a run goes wrong, and auto replay reruns failed records once a fix is in. There are no scored test cases, evaluation datasets, judge models or pass rates for an agent's decisions, and no way to compare one prompt or model against another on the same inputs.

Sourcealbato.com/embedded/pricingread 2026-11-09

Every action goes through an app's API, a webhook or a custom HTTP request inside Albato's connector engine. Albato runs no hosted browser, desktop session or remote computer for agents to drive, and for apps without an API it sets up custom webhooks rather than operating the interface. Copilot works inside Albato's own builder, and white label connect links send users to a provider's OAuth page in their own browser while the agent waits for the user to finish.

Sourcealbato.com/wiki/articles/mcp-tools-referenceread 2026-11-09