Gumloop

Connectors are the integrations an agent connects to, including Gmail, Salesforce, Slack, Notion and 150+ more, with an All tab covering Gumloop-managed and customer-added connectors and a Custom tab filtering to custom MCP servers. The connectors added determine what the agent can access. A Tool Discovery toggle controls automatic tool finding, and tools within an integration can be restricted individually.

Connectors support API key authentication as well as OAuth. The gumcp-client system skill lets skill scripts call connected integrations directly from Python using a server__tool_name convention, with a list_tools script for discovering available tools. Connector coverage keeps growing, with DocuSign, Trello, Google Docs, Google Drive and Gong among recent additions.

Sourcedocs.gumloop.com agents and skills, gumloop.com changelogread 2026-08-31

Subagents let an agent delegate to other agents instead of handling everything in one conversation. An Agent node embeds an agent inside a workflow, receiving a prompt from upstream nodes and returning output downstream, and supports Loop Mode for batch processing, where each iteration starts a new conversation unless conversation IDs are tracked.

Workflows supply branching, conditional logic and loops, and subflows are reusable modular components callable from other flows and testable independently. Triggers start workflows on schedule, webhook, event or manually, and workflows and agents can invoke each other.

Sourcedocs.gumloop.com agent_node, agents and triggers pagesread 2026-08-31

Company Brain connects knowledge sources, keeps them synced and gives agents company knowledge to search. Sources are attached whole through + Source or drilled into at the level of exact files and folders, and when a question touches internal knowledge the agent searches automatically, shown in chat as Searching Company Brain, and can open a full document for more context. Sync coverage extends to connected services, including Gong call recaps and transcripts. Knowledge is configured per agent alongside Connectors, Skills, Subagents and Triggers.

Sourcedocs.gumloop.com core concepts agents and gumloop.com changelogread 2026-08-31

Agents raise pending approval and input requests, and the right to reject has widened from the named recipient to anyone who can chat with the agent. Enterprise controls require human approval before expensive actions, alongside per-agent, per-team and organization-wide budgets and quota controls.

Per-tool authorization policies and restriction of which tools an integration can expose limit what an agent may do, connectors blocked by organization policy display as restricted with an on-the-spot access request, and skill creation can be disabled for a custom role so agents cannot save new skills. Skill Editing and Creation is a per-agent toggle.

Sourcegumloop.com enterprise page and changelog, docs.gumloop.com skillsread 2026-08-31

Gumloop says it complies with SOC 2 Type II and GDPR, and it keeps a trust center at trust.gumloop.com. Controls include role-based and attribute-based access control, per-tool authorization policies, full traceability with every call tied to a specific user, agent or service principal, a live inventory of every MCP server in use, and centrally managed secrets that never sit in plaintext locally.

Identity platform integration extends the customer's identity system of record to agents, with SCIM provisioning and per-group permission ACLs. Enterprise controls add restriction of which models teams may use, connector restriction by organization policy with request-access, and per-agent, per-team and organization-wide budgets and quotas.

Sourcegumloop.com security solutions page, enterprise page and changelogread 2026-08-31

Gumstack traces every tool call through a single logging and analytics layer, with every call tied to a specific user, agent or service principal and detailed audit trails capturing where data is flowing across the organization. It provides a live inventory of MCP servers in use and extends visibility to agents running outside Gumloop.

Organization-wide credit usage is tracked in real time with per-agent and per-team budget attribution, and agents can pull credit usage and run history reports via data export tools. Skills carry file-level version history naming the editing agent plus an activity timeline of views, script runs and edits.

Sourcegumloop.com enterprise and security pages, docs.gumloop.com skills, gumloop.com changelogread 2026-08-31

A new agent conversation starts empty, and context is kept within the chat and its sandbox files until the window fills. Search Past Conversations, on by default, lets the agent pull relevant snippets from earlier chats when they are needed. What carries forward otherwise is instruction rather than remembered fact.

Skills are reusable playbooks the agent can create and update after a correction, saved with version history, and self improve updates are written into the agent's own instructions once a person approves them. Reflections reviews recent tasks on a schedule and proposes instruction and skill changes for approval. Brain holds company knowledge from connected sources. No retention period or deletion path is stated for conversation history.

Sourcedocs.gumloop.com agent memory help, skills and reflections pagesread 2026-10-25

Gumloop runs as managed SaaS or can be deployed into the customer's own cloud as a VPC in the region of their choice, with data remaining in infrastructure the customer controls, and for private infrastructure Gumloop deploys and operates the full stack inside that cloud. Gumstack can equally run on Gumloop's infrastructure or inside the customer's own VPC. Enterprise materials list virtual private cloud deployment and MCP server hosting among enterprise entitlements. Open-weight model inference is routed through US-based providers under Zero Data Retention with no post-inference storage.

Sourcegumloop.com security solutions page, enterprise page and docs.gumloop.com ai_modelsread 2026-08-31

Skills are packaged folders containing a SKILL.md plus optional scripts, references and assets, downloadable and uploadable as .md, .zip or .skill files and shareable through Owner, Editor, Viewer and Use Only roles, where Use Only permits running a skill without visibility of its contents.

Organization Skills lets admins centrally manage a shared skill library for the whole organization, including keeping it in sync with a GitHub repository, and Gumloop links public example skill repositories and follows the Claude Skills packaging convention. Two built-in system skills, skill-creator and gumcp-client, ship with every team. A Popular Skills carousel surfaces the most-used skills in a workspace, and a template library of prebuilt flows covers common patterns.

Sourcedocs.gumloop.com skills and organization_skillsread 2026-08-31

Agents carry a Triggers section supporting scheduled and app event triggers, with schedules written in plain language and converted by an AI-powered scheduler rather than cron.

An AI Managed toggle allows the agent to create, edit and manage its own triggers during a conversation, and a Create With AI path builds custom polling triggers that combine multiple services, query connected services for exact tool names and schemas, validate in an isolated sandbox and baseline state on first run. Webhook triggers give each agent a unique URL.

Workflow triggers cover time-based runs with retry configuration plus Gmail, Slack, Teams, Airtable, Notion, Google Sheets and form events, and Outlook triggers fire on new mail or relative to calendar events. Triggers auto-disable with notification after repeated failures.

Sourcedocs.gumloop.com triggers, ai_trigger_creation and agents pages, gumloop.com changelogread 2026-08-31

Agents can use top models from every major provider, chosen in Agent Preferences at the top of the agent's configuration, with Recommended and other presets plus the full list, and new models usually arrive within a day of public release.

Every model in the agent picker supports tool calling; models without tool calling appear only in workflow AI nodes, and several open-weight models including GPT-OSS 120B, Qwen3.5 397B and Kimi K2.6 are agent-exclusive. Open-source models are accessed through US-based providers under Zero Data Retention policies with a US-provider hosted badge, with data neither used for training nor retained after inference. Enterprise controls allow restricting which models teams can use.

Sourcedocs.gumloop.com ai_models and gumloop.com enterprise pageread 2026-08-31

The Gumloop API creates agents, starts chat sessions and triggers automations programmatically, with authentication and OAuth 2.0, Python and JavaScript SDKs, and endpoints to list, create and update agents and their immutable versions, attach skills and MCP servers, create sessions, send and queue messages, upload files, resolve pending approvals, and create, update, download and delete skills.

A Gumloop CLI drives agents, sessions, MCP servers, Company Brain and skills from the shell, and the pricing page lists a Gumloop MCP among developer features. Each agent can also carry a webhook trigger URL. Custom MCP servers can also be added as connectors, which lets Gumloop call out to other tools.

Sourcedocs.gumloop.com llms.txt API reference and CLI sections, gumloop.com/pricingread 2026-09-29

Evaluations sit under each agent's Performance tab and score the agent's output. After every completed interaction an AI evaluator reads the full role-tagged transcript, including tool calls and results, and produces a grade, a pass or fail for each criterion with rationales, tags, extracted data points and optional sentiment, against quality criteria the customer configures.

The overall grade is set deterministically from which criteria failed, and a Notify grade sends notifications by email or Slack. Evaluations re-run when a conversation continues and can be backfilled after criteria change. Skill validation, sandbox-validated AI-created triggers and independently testable subflows sit beside it.

Sourcedocs.gumloop.com/core-concepts/evaluationsread 2026-09-29

The Gumloop Chrome extension turns any browser action into an automation, recording any action taken in the browser and replaying it as a repeatable automation.

It records actions on any website, replays recorded actions in Chrome and runs Gumloop automations on any webpage, for uses such as gathering and processing data from any webpage, automating tasks on websites that do not offer API access through Web Replays, and automating form filling and data entry, going beyond what traditional APIs allow. A separate web automation node handles authenticated sites and appears to be open only to some accounts.

Sourcechromewebstore.google.com/detail/gumloop/hpkeijgchoedhhdfjdjlaimanpmbnhjmread 2026-09-29