Dynamiq

Agents act across 2,100+ app integrations with 11,800+ actions, such as sending a Gmail message, posting to Slack, opening a service ticket or updating a CRM opportunity, and each Action tool's schema is generated from the catalog.

Any MCP server's tools become individual agent tools automatically, and agents query PostgreSQL, MySQL, Snowflake and Redshift directly or reach a private database through an SSH tunnel.

Credentials live server side and are referenced by connection id, either one organization connection shared across a project or a requirement that each end user of a deployed App links their own account.

Sourcegetdynamiq.ai/product/integrations.mdread 2026-10-18

Agent Builder is a drag and drop canvas and an open source Python SDK that compile to one workflow engine, so a graph of LLM, tool, sub agent and orchestrator nodes behaves the same whichever way it was built.

Decision nodes ask yes or no, choice or score questions about a record and route on the answer, rule tables hold eligibility checks and pricing grids a domain expert can edit with saved test cases, and a generator drafts nodes and edges from a plain language description. AI Coworker spawns isolated helper agents in parallel for deep research or fan out work and folds their results back in.

Sourcegetdynamiq.ai/product/agents.mdread 2026-10-18

Knowledge bases ingest uploaded files, websites crawled with depth and path filters, and Google Drive, Notion, Dropbox, SharePoint, Box and Confluence on an automatic sync schedule.

A multi file converter routes PDFs, Word, PowerPoint, images and text to the right parser, with AWS Textract, Google Document AI or an LLM for scans, and the pre processing, chunking, vectorization and storage steps are workflow nodes teams can inspect and swap across 8 vector stores and 8 embedding providers.

Metadata filters return only chunks a user or department may see, and GraphRAG adds entities and relationships alongside vector search.

Sourcegetdynamiq.ai/product/knowledge-rag.mdread 2026-10-18

Approvals hold a tool node or an agent's plan for a reviewer, who sees the exact call with its message and can edit the fields before approving it, so the step that sends an email, writes to a database or calls a payment API waits for a person.

Guardrail nodes detect names, emails, addresses and account numbers before they reach a model, classify prompt injection attempts so they can be routed away, and screen messages against content safety categories, while validators check the shape of outputs afterward. Voice agents transfer callers to labeled human destinations when a case needs a person.

Sourcegetdynamiq.ai/product/guardrails.mdread 2026-10-18

SOC 2, with the report under NDA, and CASA Tier 2, a lab verified review against the OWASP verification standard, back the platform, and Dynamiq signs a BAA for HIPAA workloads and a DPA under GDPR. Enterprise sign in runs through OpenID Connect with identity providers such as Okta or Microsoft Entra ID, limited to verified email domains, with optional two factor sign in.

Every request is checked against project membership, management needs an owner or admin role, guests see only their projects, connected system credentials sit in a HashiCorp Vault based secrets manager, and access keys are stored only as SHA 512 hashes with expiry and instant revocation.

Sourcegetdynamiq.ai/security.mdread 2026-10-18

Every run is recorded as a replayable trace, a tree of each node's input, output, status, timing, prompt and completion tokens and dollar cost.

The Monitoring tab charts requests, latency, tokens and cost by hour, day, week or month and drills from a spike into the traces behind it, workflows running in the customer's own infrastructure can send traces through a callback handler, and an AI agent combs production traces for issues and regressions for a person to check.

Traces export as JSON, prompt and response content can be stripped immediately or after set hours, and voice calls show a per turn latency waterfall.

Sourcegetdynamiq.ai/product/observability.mdread 2026-10-18

With memory on, an Agent node keeps each conversation scoped by user and session and replays the most recent messages, the most relevant ones, or both, at the start of the next run. Long term memory gives the agent tools to save, recall and update durable facts about each person, which carry into every new conversation. In AI Coworker, those notes persist across conversations and people can view and correct what it remembers on a Memory page, while conversations, files and scheduled tasks stay private to their owner.

Sourcegetdynamiq.ai/product/agents.mdread 2026-10-18

Four placements run the same platform: Dynamiq Cloud as a managed service, the customer's own cloud account on AWS, Azure, GCP, IBM Cloud or Red Hat OpenShift installed from the official Helm chart, any Kubernetes 1.32+ cluster in the customer's data center backed by its own database, message bus and object storage, and fully isolated environments with no outside access delivered with Dynamiq engineers. When self hosted, documents, knowledge bases, traces and credentials stay in the customer's storage, outbound calls go only to the configured model and tool providers, and secrets can come from the customer's own store such as AWS Secrets Manager.

Sourcegetdynamiq.ai/deployment/self-hosted.mdread 2026-10-18

AI Coworker comes ready to use as an agent that plans, browses, writes code and works connected apps without any build step, and it loads packaged skills, official ones or skills a team writes and shares, which carry over to workflow agents. The Agent Builder generator drafts a workflow from a plain language description, and use case pages describe agents for customer service, KYC, claims, compliance, accounts payable and research. A gallery of ready made workflow templates to install is not offered, so most agents start from the generator or a blank canvas.

Sourcegetdynamiq.ai/product/chat.mdread 2026-10-18

Deployed Apps answer HTTP calls synchronously, by server sent event stream or asynchronously with a callback URL, embed as a chat widget or hosted chat link, and run on a cron schedule, at a one off time or from any of 3,300+ app triggers such as a new Slack message or an incoming email. AI Coworker answers in Slack, where it reads the thread and replies with files, Microsoft Teams and Telegram, and runs prompts daily, weekly or monthly. Voice Agents take inbound calls over SIP on numbers the customer brings, WebRTC from a browser or app, or a WebSocket from the customer's backend.

Sourcegetdynamiq.ai/product/deployments.mdread 2026-10-18

The SDK supports 29 model providers, from OpenAI and Anthropic to self hosted Ollama, with every provider node sharing the same parameters, streaming and tool calling shape.

The AI Gateway is one OpenAI compatible endpoint for the most used models, holding credentials server side behind a single Dynamiq key, switching providers by model slug and routing to a fallback, a self hosted open model for traffic with personal data or a regional deployment.

Teams can host and fine tune open models with LoRA adapters, Anthropic and Bedrock calls reuse cached prompt prefixes, and AI Coworker picks the model per conversation.

Sourcegetdynamiq.ai/product/models.mdread 2026-10-18

The dynamiq Python SDK is open source under Apache 2.0, free for commercial use, and builds the same workflows the canvas does, so engineers can version agents as code and deploy them to the platform.

Every deployed App exposes an HTTP API with sync, streaming and async modes, a management API covers the platform, the AI Gateway offers an OpenAI compatible endpoint, and voice agents accept client tools that the customer's app answers.

MCP servers plug in as tool sources, organizations register their own MCP servers once for AI Coworker, and tracing callbacks bring runs from outside infrastructure into the same view.

Sourcegetdynamiq.ai/product/agents.mdread 2026-10-18

Evaluations score agent outputs with an LLM as a judge rubric, predefined RAG evaluators such as faithfulness, or custom Python functions, over versioned datasets built by hand, uploaded or captured from production traces with one click, so a real failure becomes a permanent regression test. A candidate workflow version can be rerun against a released dataset and compared with production scores, a metric attached to a deployed App scores a sampled share of live traffic without extra latency, and voice agents face scripted caller simulations graded on pass rate, accuracy and experience before real callers reach them.

Sourcegetdynamiq.ai/product/evaluations.mdread 2026-10-18

Each AI Coworker conversation gets its own isolated Linux sandbox, and inside it the agent drives a browser through websites on its own, handing control back to the person when a login is needed and then carrying on.

The same sandbox runs shell commands, installs packages, writes files and serves live previews of sites the agent builds, viewable at desktop or mobile width. That browser lives in Coworker.

Workflow agents built on the canvas reach outside systems through integrations, MCP servers and SQL, and voice agents can send keypad tones to walk another company's phone menu, which is telephony and involves no screen.

Sourcegetdynamiq.ai/product/chat.mdread 2026-10-18