Viktor

Viktor connects to 3,200+ tools, mostly by narrowly scoped OAuth, and does real work in them. The product page says customers connect "3,200+ tools, from the CRM to the ad accounts" (Google Ads, Google Drive, Stripe, HubSpot, Notion, Linear, Asana, ClickUp, Cursor, GitHub), and Viktor works in them and posts finished work back; the site publishes per-integration pages (HubSpot, Salesforce, Jira, QuickBooks, Meta Ads and others).

The security page says "every major tool connects via OAuth with the narrowest scopes that get the job done", API keys sit in an AES-256 vault injected at execution time by a backend tool gateway, access-logged and rotatable, and admins decide which integrations are connected and who can use them. Actions are real (syncing MQLs, drafting invoices, building apps), and credentials are narrowly scoped, rotatable and revocable in one click.

The docs home says 9,000+ integrations while the product page says 3,200+.

Sourceviktor.com/product, /security; read; viktor.com/securityread 2026-09-16

Multi step jobs across several tools run to completion, mixing fixed tool operations with the agent's own judgment.

The product page shows Viktor sorting overnight support tickets, "answering where he can" while "the ugly ones are escalated by name, with the customer's history attached"; pulling every open deal from the CRM and working out what moved and which deals deserve a call; and building an internal order tracker on Shopify and Stripe that he keeps running and uses to flag late orders into a channel.

The No API page shows a job that pulls overdue invoices in QuickBooks, drafts reminders for the oldest, posts a spreadsheet and holds the send. Jobs combine fixed tool operations (pull, post, draft) with agent judgment, and routing each ticket to an answer or a named escalation is a branch.

Retry and fallback configuration and versioning are not documented, and orchestration is shown through examples rather than a workflow reference.

Sourceviktor.com/product, /no-api; readread 2026-09-16

Answers come from the tools and team threads a workspace connects. The product page shows Viktor answering "what did we decide on the Q3 supplier switch?" from a June thread with a summary attached, working inside the tools a team connects (Google Drive, Notion, HubSpot, GitHub and 3,200+ others) and pulling data from them for reports.

No index of connected content, source refresh or citation surface is documented; context is assembled per run. Skills, Viktor's persistent memory, are a separate store.

Sourceviktor.com/product; readread 2026-09-16

Irreversible actions and new scheduled tasks wait for approval, and admins choose which tools need a yes. The product page says Viktor "asks before doing anything you can't undo": emailing a customer list or touching what people pay "waits in the thread until someone presses Approve", with Approve and Reject shown on a held send; a proposed scheduled task "stays paused until someone approves it"; and a thin brief gets questions while a wrong-looking number gets flagged.

The security page says money moves, code pushes and customer emails wait for explicit approval in Slack, and "admins put high-risk tools behind human approval"; the No API page says "you choose which actions need a yes in Slack or Teams before they happen in the tool". Approvals sit at tool and action level, arrive in Slack or Teams, and vary by risk.

Why a checkpoint fired is shown by example rather than documented as a feature.

Sourceviktor.com/product, /security, /no-api; readread 2026-09-16

A SOC 2 Type 1 report, SAML single sign on and admin control over integrations are all documented. The security page's compliance table lists SOC 2 Type 1 as certified, "independent attestation that our security controls operate as designed", with the report available under NDA; Type II and ISO 27001 are listed as in progress, and CASA Tier 3 is shown.

For access, SAML SSO works across Okta (inside Slack), Entra ID, Google Workspace, OneLogin and any SAML 2.0 IdP; admins decide which integrations are connected, who can use them and at what level; and admins can disconnect any integration, pause any user or kill a running task in one click. Viktor and its model providers are bound by no-training commitments.

SCIM provisioning is not named, and the SOC 2 report sits behind an NDA.

Sourceviktor.com/security; readread 2026-09-16

Usage is reported per run, changes are posted back and credential access is logged. The product page says anyone can ask in the channel what the week used and Viktor "answers from your live usage", with the web app keeping the full breakdown by run count and cost driver, and admins get credit-burn warnings and forecasts. The No API page says Viktor "posts back what changed" after working in a tool, and the security page says vaulted credential access is "access-logged".

Inspecting prompts, tool calls, retrieved knowledge and outputs step by step, and export to SIEM, are not documented.

Sourceviktor.com/product, /no-api, /security; readread 2026-09-16

Viktor keeps a persistent, workspace-scoped memory of written skills that a team can read, correct, export and delete. The security page names "Skills, the persistent memory": "memory is scoped to your workspace, encrypted at rest, never used to train models, and fully exportable or deletable on request", with skills and memory walled off per workspace.

The product page says Viktor turns a screen recording "into a skill: a written procedure you can open and correct", that "what he learned sits on paper where you can read it", and that he "keeps what he learns" so week three is better than week one. Scope is the workspace, lifetime is until export or deletion, and skills are separate from the business record.

Per user scoping and handling of stale memory are not described.

Sourceviktor.com/security, /product; readread 2026-09-16

Hosting is in the US by default, with EU data residency on Enterprise contracts. The security page, under data handling, says "Data residency options: US-hosted by default. EU data residency available on Enterprise contracts", and each workspace is fully isolated, with skills, integrations and memory walled off per workspace; the FAQ asks "Where is data hosted, and can we choose a region?".

EU residency comes only with an Enterprise contract, and no self hosted or customer cloud deployment is offered.

Sourceviktor.com/security; readread 2026-09-16

Viktor is one configurable AI employee, with no catalog of packaged agents or templates. The use-case page ("One AI employee.

Every function.") lists example jobs by function, such as pausing ad sets by CPA threshold, weekly P&L summaries from Stripe and Shopify, and branded client recaps, each shown as a Slack request to the same agent; it offers no catalog entry, template or packaged agent to install. The product page shows Viktor proposing scheduled tasks drafted from a customer's own recording.

Example requests to one agent are not packaged assets, and nothing on the page stands as a separate product. Skills a customer records become memory, not packaged templates.

Sourceviktor.com/use-case, /product; readread 2026-09-16

Standing work runs on schedules with nobody asking, and people reach the agent in Slack, Teams and through its API. The product page shows standing work such as "Mondays, 9:00 #marketing ad spend recap", "Fridays, 16:00 #sales pipeline update" and "Every morning, 7:00 #support inbox triage", and says "the Monday report is in the channel before you open your laptop". Viktor proposes scheduled tasks from recorded skills, which run once approved; otherwise users message him in Slack or Microsoft Teams, and the Public API starts tasks from other applications.

Event triggers, such as a new ticket or record starting a run, are not documented. The approval before a schedule's first run is an oversight gate, not a limit on the trigger.

Sourceviktor.com/product; readread 2026-09-16

Customers choose how hard the agent thinks, and each setting maps to a named model. Under "You decide how hard he thinks", the product page offers three settings: Smart (Claude Opus 5, low reasoning, base rate), Balanced (Claude Sonnet 5, medium reasoning, about 50% cheaper) and Ultra (Claude Fable 5, high reasoning, about 2x cost), and says Viktor picks up better models as they ship. The security page says "inference runs on OpenAI, Anthropic, and Google", each on the public sub-processor list with a no-training agreement.

The choice is among Viktor's three tiers, not an arbitrary or bring your own model, and whether the setting applies per task, per workspace or per scheduled job is not stated.

Sourceviktor.com/product, /security; readread 2026-09-16

A public API, a served MCP server, and endpoints compatible with OpenAI and Anthropic are documented for the platform itself.

The developer docs say the Viktor Public API lets developers "create Viktor threads, run tasks, continue conversations, and download result files from your own applications"; the Viktor MCP Server lets an MCP-compatible agent connect "with streamable HTTP and a scoped Viktor API key"; and an OpenAI- and Anthropic-compatible API drives Viktor agent runs with the OpenAI or Anthropic SDKs. A public API with SDK-compatible endpoints and a served MCP server let Viktor fit into other agent systems.

Reference pages for each (/docs/public-api, /docs/mcp-server, /docs/chat-completions) are linked from the docs home.

Sourceviktor.com/docs; readread 2026-09-16

No way to test a change to skills, instructions or scheduled tasks, before or after it goes live, is documented. New scheduled tasks are held paused for human approval and irreversible actions wait for Approve or Reject; Viktor questions thin briefs and flags numbers that look wrong before reporting them; and usage and cost are reported per run. None of these tests a change: there is no simulation, dry run, test case or quality score.

Sourceviktor.com/product, /security, /no-api, /docs; readread 2026-09-16

Viktor signs into tools that have no API with its own seat, and works on their screens from its own cloud computer.

The No API page says "if a person can use the software, Viktor can": customers create a seat for Viktor in a tool with its own email and permissions, and "Viktor signs in from his own computer in the cloud and works on the same screens your team uses", for example logging into Greenhouse to move candidates between stages, or working in QuickBooks to pull overdue invoices and draft reminders, without an API or developer key.

Because the agent works third party web applications through their user interface, how often it breaks when UI elements change is a live question.

Resilience to UI changes and the list of supported applications are not documented.

Sourceviktor.com/no-api; readread 2026-09-16