GPTBots

A LoopAgent calls whatever is bound to it: official tools such as Google search, custom API tools built visually, MCP servers, published workflows that become tools with their start node inputs as parameters, data tables it can query and write, skills, forms and human service.

Credentials such as URLs, API keys and auth headers stay on the platform and are injected at execution, so the model never sees them, and with more than 30 tools the agent switches to lazy loading to keep token use down. A failed call shows on the tool card while the agent retries or reports it, and Zapier and n8n connect workflows to outside automation.

Sourcegptbots.ai/docs/tutorial/bot/loop-agent/toolsread 2026-10-18

Three agent types cover the range from fixed to autonomous. FlowAgents and the Flow 2.0 workflow builder wire LLM, intent recognition, condition, loop, batch processing, code, HTTP, tool, knowledge retrieval, data table and variable nodes on a drag and drop canvas; LoopAgents let the model choose each next action within a capped think, act and observe loop; and MultiAgent runs a planner that turns a request into a task list, generates actions for each task, assigns them to executor agents, has a reviewer check each result against custom standards and updates the plan as it goes, with retry limits per task.

Sourcegptbots.ai/docs/tutorial/bot/multi-agent/overviewread 2026-10-18

The knowledge base takes unstructured files in PDF, DOC, Markdown and text, structured CSV, XLS and XLSX tables and question and answer pairs, with Google Drive and Notion as sources, and retrieves through hybrid semantic and keyword search with multi channel recall, an optional rerank step and citations shown with answers.

Documents carry metadata fields, knowledge bases have their own permissions, and APIs add documents, spreadsheets, text chunks and Q&A pairs, sync a knowledge base, re embed failed documents and run vector similarity matching. Text2SQL lets agents query databases and turn results into charts.

Sourcegptbots.ai/docs/tutorial/knowledge-base/overviewread 2026-10-18

Workspace tools carry permission levels: silent for reads, a permission card with an always allow option for commands, file writes and the browser, and a card on every call for high risk actions, with deletes and commands such as rm, sudo or dd always forced to confirmation and a rejected action stopping the agent.

MultiAgent asks the user to confirm its task list before running and lets a person take over the Computer Use desktop. Conversations hand off to human agents in Intercom, LiveDesk, LiveChat, Zoho SalesIQ or Sobot during set service hours with recent rounds attached, and moderation can block unsafe input or output.

Sourcegptbots.ai/docs/workspace/app-settings/runtime-securityread 2026-10-18

Organization roles set feature permissions, which agents a member can operate, Workspace rights and a credit spending limit per role, and knowledge bases and databases have their own permissions. SAML 2.0 single sign on works with Microsoft Azure, Okta, OneLogin, Google and AWS Cognito. GPTBots holds ISO 27001 for information security, ISO 27701 for privacy information management and a SOC 2 Type II report.

PII anonymization through Microsoft Presidio replaces names and other entities before text reaches a model and restores them afterward, with custom entities defined by regex, word lists and context, and content moderation through OpenAI Moderation, keyword lists or a custom API screens text and images.

Sourcegptbots.ai/docs/tutorial/organization/member-and-roleread 2026-10-18

Agent logs keep chat records with question classification and sentiment, and a log tree API returns the full execution trace of a conversation.

The Workspace log audit lists each conversation's source channel, status, processing node, tools used with call counts, message count and cumulative tokens, plus AI search queries and A2A agent interactions, with permissions per tab and logs isolated by organization.

Analytics APIs report credit use per conversation and workflow run, human handoff rates, peak requests per minute and recent agent activity, and alarm notifications warn administrators about business events.

Sourcegptbots.ai/docs/workspace/space-manage/log-reviewread 2026-10-18

Memory runs on two layers. Short term memory keeps the last N rounds of a conversation, 30 by default, and a separate full context setting carries the tool calls, parameters and results of the most recent rounds so the agent remembers what it just did. Key events are cross session, user level long term memory that the model creates and updates during conversations, much like a ticket that follows the customer, and they appear in the user list for each end user. In the Workspace, the assistant remembers personal preferences such as a favored programming language across sessions.

Sourcegptbots.ai/docs/tutorial/bot/loop-agent/memoryread 2026-10-18

Private deployment puts the platform in the customer's own environment, cloud native or on premises, for banks, financial institutions and government bodies that need full data control, with its own brand, domain and account system, a model management layer for private, fine tuned and commercial models, and one or two product updates a year.

GPTBots runs the rollout from requirements and environment preparation through delivery, with hardware advice for on premises inference. The Workspace assistant can also run tasks on the employee's own device or in a local sandbox restricted to chosen folders.

Sourcegptbots.ai/company/private-deploymentread 2026-10-18

A template gallery offers ready agents in customer support, AI SDR, enterprise search, data insights and productivity categories with one click deployment, and packaged solutions cover customer support, lead generation, enterprise search and data analytics. MultiAgent ships with preconfigured agents such as an AI coder, Browser Use, Computer Use and online search, plus roles across development, product, testing, data and marketing. The Workspace marketplace adds installable apps such as Workpilot and Live Speechly, and agents export and import as .bot files.

Sourcegptbots.ai/templatesread 2026-10-18

Agents publish to a website bubble or iframe widget and a share link, WhatsApp through Meta or EngageLab, Facebook Messenger, Instagram, Slack, Microsoft Teams, Telegram, Discord, DingTalk, WeChat customer service, LiveChat, LiveDesk, Sobot and Zoho SalesIQ, and to Zapier, with REST and webhook APIs for anything else. In the Workspace, scheduled tasks run a saved prompt daily, weekly or on custom days, locally or in a sandbox, pinned to a model and sent to Telegram, WhatsApp, Discord, Slack, Teams, WeChat or DingTalk when done, and employees can mention the agent inside Telegram, DingTalk or Lark groups.

Sourcegptbots.ai/docs/tutorial/bot/integration/overviewread 2026-10-18

Agents switch among OpenAI GPT 5.1, GPT 5, GPT 4o and GPT 4o mini, Anthropic Claude including Sonnet 4.6 thinking, Google Gemini, DeepSeek and other models, priced in credits per thousand tokens, on GPTBots managed keys or the organization's own API keys. Separate models handle speech recognition, text to speech, embeddings and reranking, the MultiAgent planner takes a model chosen for speed and reasoning, and Workspace scheduled tasks can pin a model or follow the global default. Private deployments add LLM scheduling and load balancing across private, fine tuned and commercial models.

Sourcegptbots.ai/docs/getting-started/model-listread 2026-10-18

REST APIs reach almost every surface: account APIs import agents and workflows and create tools, MCP servers and skill packages; conversation APIs create conversations, send messages, share links, query the execution log tree and manage agent versions; and knowledge base, database and analytics APIs handle documents, tables and usage. Webhooks receive agent replies and human handoff messages, extension apps sign users in through an SDK, and iOS and Android SDKs embed agents in mobile apps. An agent skill lets Codex or Claude generate a .bot configuration and push it to a test mode agent.

Sourcegptbots.ai/docs/api-reference/overviewread 2026-10-18

Agents created in test mode can be updated and released through the API, versions can be listed, released and rolled back, and debug preview runs with a clean context. A quality API labels responses as fully or partially resolved, a feedback API records user ratings, and logs classify questions by topic and sentiment for review. Saved test sets, automated scoring of answers against expected results or side by side comparison of versions are not offered.

Sourcegptbots.ai/docs/api-reference/conversation-api/quality/qualityread 2026-10-18

MultiAgent includes a Browser Use agent that carries out tasks in a sandboxed browser and works even with models that cannot read images, and a Computer Use agent that runs in a sandboxed Linux environment with desktop apps, the file system and system commands, with iteration limits and a takeover mode that lets a person operate the computer remotely when the agent gets stuck. In the Workspace, the assistant can open Word, Excel, a browser and a shell on the employee's device or in a sandbox, behind tool permission prompts.

Sourcegptbots.ai/docs/tutorial/bot/multi-agent/overviewread 2026-10-18