Serval

The integration catalog runs to 164 documentation pages, listed as its own index at docs.serval.com/llms.txt, with a page per connector rather than a logo on a grid. Named systems span every category the product touches: ticketing with ServiceNow, Jira, Zendesk and Freshservice, bidirectional so a customer can keep existing tooling during migration; identity with Okta, Microsoft Entra, LDAP and Active Directory, plus SailPoint for governance and RSA SecurID for MFA; endpoints with Jamf and Tanium; HR with Workday; knowledge with Notion, Confluence, Google Drive and Slab; collaboration with Slack, Microsoft Teams and Google Workspace; and finance and operations with Coupa, Salesforce, LaunchDarkly, Vanta and Presidio.

The connector pages show authenticated action, not data mirroring. The RSA SecurID page describes Serval storing the client key encrypted and "attaching it automatically to every call it makes on your behalf", with workflows driving a full MFA flow: starting, verifying, checking and canceling attempts on demand. The Presidio page describes Serval exchanging credentials for a short lived JWT and reading asset records "on your behalf whenever a Serval workflow runs", with the regional endpoint chosen by the workflow. Both happen at execution rather than by background sync.

Provisioning is the heaviest write path: automated provisioning and deprovisioning via SCIM, direct API or custom workflows, with just in time access grants revoked automatically when the window ends. Access revocation via API shipped in February 2026. The agent can see the connectors too: the RSA page says "Serval's AI knows the endpoint catalog and can help you build workflows against it". Connector maturity is labeled, with RSA marked Beta in the integrations UI.

Sourcedocs.serval.com RSA SecurID and Presidio integration pages, with the integrations indexread 2026-09-12

Composition is a first class feature. Child workflow composition shipped in February 2026, so a workflow can call another as a step; workflow journeys can be viewed as dependency graphs since August 2026; and workflows can be organized into folders, shared across teams, tagged, searched semantically and versioned with diffs.

The runtime is named in the architecture. Temporal is listed among the required datastores as the workflow orchestration component: a durable execution engine, so a run survives process restarts and resumes. That is a concrete answer to how long running, multi step work interrupted by people is held together. Pauses and branches are documented: pause workflows for approval shipped in April 2026, conditional forms in Slack in March 2026, criteria based assignment rules in February 2026, a default triage team for ambiguous requests in May 2026, and ITIL aligned change management in June 2026.

Authoring is conversational and compiles to code. The Automation Agent turns natural language descriptions into TypeScript, so automations build themselves from a description. Serval markets deterministic workflows rather than execution the model improvises: the model writes the workflow, and the workflow then runs the same way every time. Workflows and skills are managed as code in Git with version diffs.

The participants are distinct: the employee raising a request, three named agents with different jobs, approvers who can be delegated to or overridden by a team manager, assignment groups, and external ticket subscribers. Work spans ticketing, just in time access, asset management, onboarding and offboarding, compliance workflows and campaigns, with Security, IT, HR, Finance and Operations each able to run isolated environments on the same platform.

Sourcedocs.serval.com release notes index and self-hosting prerequisitesread 2026-09-12

Serval now owns its knowledge store rather than only syncing one. A native knowledge base shipped in beta in June 2026, and the release history shows it built out as a maintained structure: knowledge base management updates in November 2025, knowledge tagging within guidance, lexical and keyword search added alongside semantic retrieval in March 2026, cross team KB sharing in August 2026, and a tag catalog with owners set for teams, resources and roles in June 2026. Ownership, tagging and sharing are what a curated corpus looks like.

The synced estate runs both ways. Answers ground in help docs built in Serval or synced from Notion, Google Drive, Confluence and Slab, and since May 2026 "reviewers can publish approved knowledge updates into Notion, Slab, Confluence, Freshservice, and ServiceNow". Corrections flow back out to the systems of record, so the corpus is maintained rather than merely read.

Retrieval respects permissions at read time: knowledge source permission enforcement shipped in January 2026, so an employee cannot be answered from a document they are not entitled to see. The corpus goes beyond documents. The Help Desk Agent gained web search in March 2026; tickets, assets and access history are queryable alongside; and the Insights Agent analyzes ticket patterns. Serval's asset estate tracks hardware, software licenses, SaaS subscriptions and cloud costs in one schema the agent reasons over.

Sourcedocs.serval.com release notes indexread 2026-09-12

The approval gate is a built surface with its own release history. "Pause workflows for approval" shipped in April 2026: execution halts mid flow and waits for a person. A redesigned approval flow builder shipped in August 2026, so customers author the gates.

Around them: custom approval workflows in December 2025, natural language workflow approvals in January 2026, approval delegation in March 2026, team managers able to override approvals from a ticket in July 2026, and an approvals API in April 2026 for custom follow up logic. Multi step approvals apply to workflows and access requests alike.

The gate sits on the consequential action. Access requests route through approval before a grant is made, just in time access is time bound and revoked automatically when the window ends, and the access history export records the approver and the justification against every grant.

Guardrails work where approval is not requested. Serval markets deterministic workflows with security layers: "RBAC, approval gates, API scope limits, air gap between help desk and builder." That last one is real containment: the agent that talks to employees cannot edit workflows or integration scopes, so a compromised conversation cannot widen the agent's own authority. Skills and runbooks hold agent behavior to the team's written procedures, with a diff view for reviewing changes the AI suggests before they take effect.

Sourcedocs.serval.com release notes index and serval.com/insights buyer's guideread 2026-09-12

SOC 2 Type II is held, with the wording varying by page. Serval's compliance update says "by attaining SOC 2 Type 2 compliance, our internal controls and processes have been thoroughly vetted, coupled with our adherence to HIPAA and GDPR regulations".

The regulated industries page is firmer: "Serval holds SOC 2 Type II and documents HIPAA-oriented deployment and data controls." The homepage is looser: "complies with SOC 2, HIPAA, and GDPR frameworks".

Serval tells buyers to "ask vendors for the current Type II attestation report, not a marketing badge" and points to its own Trust Center, alongside a written BAA for HIPAA workloads and a DPA for GDPR.

The customer facing controls are deep and centralize access management: "SAML and SCIM support enable seamless SSO and user provisioning", role based access control for least privilege permissions, customizable API scopes for app integrations, and multi step approvals for workflows and access requests. The release history shows the surface growing: custom roles, scoped API keys, knowledge source permission enforcement, ticket visibility groups for access controlled routing, a read only agent role, default access profiles and app owners, skills gated by access profile, and sharable access reviews.

A separation of duties control is named: an air gap between help desk and builder, shown as the help desk agent being unable to edit workflows or integration scopes. The agent that talks to employees cannot change what agents are allowed to do.

Sourceserval.com/updates/soc2, /insights/hipaa-compliant-itsm and the serval.com platform pageread 2026-09-12

Every workflow run is logged step by step: "Serval logs every workflow run with inputs, outputs, status, and duration." That answers what the agent did, not just why it answered as it did.

The log can be exported in a stated format. Serval offers a "sample workflow run export with per-step detail and timestamps" and an "access history export with approver, justification, start/end, revocation reason", and tells buyers to "confirm logs include per-step inputs and outputs, precise timestamps, and CSV/JSON export for GRC integration".

An access record carrying the approver and the justification is an audit trail of a consequential agent action with the human decision attached. SIEM integration carries it into the customer's security estate, and Serval frames the point as evidence "generated when work happens, not reconstructed before an audit".

Run debugging and version history sit alongside: agent powered run debugging shipped in February 2026, version diffs for workflows and skills in July 2026, a guidance diff view for reviewing AI suggestions in January 2026, automation suggestion history in December 2025, and workflow journeys viewable as dependency graphs in August 2026. A reader can see what ran, what changed and what the agent proposed.

Dashboards on request volume, time to resolution and SLA compliance measure the help desk operation; AI resolution rates read on what the agent itself did, and the Insights Agent reports on ticket patterns and automation opportunities. Analytics dashboards went to public beta in June 2026, with a refreshed home in September 2026.

Sourceserval.com/insights SOC 2 governance page, with docs.serval.com release notesread 2026-09-12

Per team guidance, knowledge sources and configuration persist in Serval, but none of these is a memory that learns per user over time, and no agent memory is documented. Guidance, renamed Skills in May 2026, is a set of written procedures and runbooks the customer authors so agents follow internal practice, with a diff view for suggested changes and version history.

It persists because a person wrote it down, not because the agent retained anything. The native knowledge base and the corpora synced from Notion, Confluence, Google Drive and Slab are the knowledge layer, and workflows, skills and settings managed as code in Git are authored artifacts.

The runtime architecture rules it out. The self hosting prerequisites list the datastores Serval needs and label each: "PostgreSQL as primary application database", "Redis for caching and ephemeral state" and "Temporal for workflow orchestration". Redis is named as ephemeral by the vendor. Temporal keeps a workflow run alive through restarts, but that is state within one execution. No component in the stack holds what an agent learned.

The documentation and release notes show no named memory layer, no retained preference the agents reapply for a user or requester, no decision trace carried between tickets, no stated scope or lifetime for anything held, and no customer surface to inspect, edit, export or purge agent state apart from tickets, skills and knowledge. Ticket history is a record of work.

Sourcedocs.serval.com self-hosting prerequisites and release notes indexread 2026-09-12

Four delivery models are documented. Fully managed Serval Cloud has "no infrastructure to manage", automatic updates and a 99.9% uptime SLA. Hybrid Self-Hosted has the customer run "only the Serval worker" inside its own infrastructure while Serval operates the backend, control plane, web app and database, so the worker "is the only Serval component that ever touches your internal network or handles your integration credentials". Fully self hosted comes in two operating modes: Managed in Your AWS, and Self-Managed on Your Kubernetes.

The fully self hosted model is single tenant in the customer's own account, and the boundary is explicit: "You bring an AWS account you own", and "the application runtime and all customer data remain in that account, and your AWS credentials never leave your environment." Either the customer's team or Serval operates it, and the deployment is identical either way.

The detail is the kind that only exists where the option is real: a deployment bundle of Terraform, Kubernetes and Helm configuration and setup scripts; an AWS stack of EKS, Aurora PostgreSQL, ElastiCache, OpenSearch, S3 and Secrets Manager; the option to point at the customer's own PostgreSQL, Redis, Temporal, OpenSearch or S3 compatible storage, with the upgrade cost of doing so spelled out; DNS requirements down to the app and API hostnames; and private network patterns named as Site-to-Site VPN, Direct Connect, Transit Gateway and private DNS resolution.

The cloud product offers a region choice too. A June 2026 release states that organizations can opt for EU data residency, and August 2026 adds Microsoft GCC High support for US government workloads. Installation and day two operations sit in a password protected area shared with customers and evaluators, while everything needed to evaluate the models is public.

Sourcedocs.serval.com self-hosting section and release notesread 2026-09-12

Three named agents ship, each doing a different job: a Help Desk Agent handling intake and resolution across Slack, Teams, email, phone and portal; an Automation Agent turning natural language descriptions into TypeScript; and an Insights Agent doing pattern detection and suggesting automations. Remove the Insights Agent and the help desk still resolves tickets; remove the Automation Agent and the other two still work. They are separate agents, not stages of one pipeline.

A browsable, installable catalog sits alongside the agents. The release history records installable workflow bundles in April 2026, installable workflow browsing in January 2026, and packs for specific estates: new installable workflows for LDAP/AD, Okta and Google Workspace in April 2026, with Google Workspace workflows installable later that month.

Workflow backed catalog items and service catalog ingestion shipped in February 2026. An installable Okta workflow does work when selected, and removing it leaves the others whole. Workflows can be organized into folders, shared across teams, searched semantically and tagged.

A native knowledge base and Skills add reusable content, though the customer writes those. The catalog comes from Serval, with no third party or community contribution path and no marketplace of externally built agents.

Sourceserval.com/insights buyer's guide, first party about Serval, with docs.serval.com release notesread 2026-09-12

Event triggering has shipped and has a release history: event triggered workflows in December 2025, extended through 2026; webhook triggered workflows in November 2025, which let any external system act as an event source; Slack reaction triggers in February 2026; SLA workflow triggers in April 2026; triggering workflows when a ticket is assigned in July 2026; and campaigns triggered directly from a workflow in June 2026. Child workflow composition lets one flow start another. Tickets, access requests and onboarding events all start work.

Schedules are a separate mechanism: recurring campaigns in April 2026, assignment schedules in March 2026, SLAs redesigned with business hours and pause conditions in February 2026, and automatic ticket bumps for unresponsive requesters in July 2026, a time based trigger acting without a person.

The channel set is wide, and the vendor lists it: the Help Desk Agent covers "intake and resolution across Slack, Teams, email, phone, portal". Phone is a live channel many products lack. The release history adds an embedded widget with a live agent in September 2026, multiple email intake addresses per team in August 2026, Microsoft Teams notifications and reply routing improvements, conditional forms in Slack, dynamically rendered forms, a Serval Inbox consolidating notifications, and white label Slack bot customization. The UI ships in nine languages with on demand ticket translation.

Sourcedocs.serval.com release notes index and serval.com/insights buyer's guideread 2026-09-12

Serval names two model providers and splits the work between them, but the customer does not choose. The self hosting prerequisites require the customer to "provide your own keys for the AI providers Serval uses at runtime" and list them: an Anthropic API key as the "primary provider for AI agents and help desk functionality", and an OpenAI API key "used for embeddings, image description, and specific model tasks".

The customer holds the credentials. The Managed in Your AWS model lists, under what you own, "AI provider keys and any customer-owned credentials supplied during onboarding", and a release note adds an in product surface to manage Anthropic usage. So the buyer owns the provider relationship, sees the spend and controls the data path to the model.

Bringing your own key is control over credentials, cost and data path, not model selection. Serval decides which provider handles which task, and nothing offers a model picker, a per agent model assignment, a version pin, or a way to substitute another provider for either role. The key arrangement is documented for the self hosted and hybrid models; nothing says a Serval Cloud customer supplies keys.

Sourcedocs.serval.com self-hosting overview and prerequisitesread 2026-09-12

Serval can be called from outside through a large, documented API. The documentation index published at docs.serval.com/llms.txt lists three sections: Documentation at 142 pages, Integrations at 164 pages, and an API reference at 252 pages. Two OpenAPI documents are listed by name, auth-openapi-spec.yaml and openapi-spec-v2.yaml, served from the API reference section, so the surface is a published contract rather than an incidental one.

The authentication model is named and has its own release history: scoped API keys shipped in August 2026, and customizable API scopes for app integrations are a stated platform property, so access is granted per integration rather than all or nothing. Named endpoints exist for the actions that matter: an approvals API for custom follow up logic shipped in April 2026, and access revocation via API and workflows in February 2026. Both let an outside caller drive a consequential action inside Serval.

Two more surfaces sit beyond the REST API: a CLI, with access requests via the CLI shipped in March 2026, and user MCP, shipped in the same release, which puts Serval inside the customer's own assistant. Webhook triggered workflows cover the event side, and workflows, guidance and configuration are managed as code in Git repositories.

Sourcedocs.serval.com llms.txt index, OpenAPI spec listings and release notesread 2026-09-12

An agent evaluation feature exists, but it is in beta and its output is not described. The release notes for July 31, 2026 record "Help Desk Agent evaluations in beta": a harness aimed at the agent's own behavior rather than at the customer's ticket queue. What an evaluation produces is not stated: no score, pass rate, judge verdict or comparison format.

The surrounding tools are real. Workflows and skills are managed as code in Git, with version diffs for workflows and skills shipped in July 2026, so a change can be reviewed before it lands. Agent powered run debugging shipped in February 2026, so a failed run can be inspected. A guidance diff view for reviewing AI suggestions shipped in January 2026, gating the agent's own proposed changes to its instructions. Serval markets deterministic workflows, so behavior repeats between runs. Version control plus a review step is a quality gate, not a test with a result.

Serval's commercial claim is a guaranteed automation rate, citing customer automation percentages. An automation rate measures how much of the customer's ticket volume the agent cleared: an outcome on the operation and a commercial guarantee, not an evaluation of the agent.

Sourcedocs.serval.com release notes indexread 2026-09-12

Integrations and APIs are how Serval acts, not driving a browser or a computer interface. The mechanism is documented connector by connector, and it is always an API. The RSA SecurID page describes Serval storing a client key and "attaching it automatically to every call it makes on your behalf"; the Presidio page describes exchanging credentials for a short lived JWT and sending it as a bearer token on every request.

Provisioning runs via SCIM, direct API or custom workflows. Across 164 documented connectors the pattern is authenticated service to service calls, and the self hosted worker exists to reach internal APIs, LDAP and Active Directory over the network, not to operate a desktop. Serval markets deterministic workflows compiled to TypeScript, which screen driving automation cannot be.

Endpoint management looks like remote control and is not. Serval integrates Jamf and Tanium and runs device actions across managed endpoints, and its governance material refers to every "grant, device action, and workflow run" being exportable. Those actions go through the MDM and endpoint platforms' own APIs. The CLI, which can raise access requests since March 2026, is a way for people to reach Serval, not Serval operating anything.

No hosted or local browser, desktop session, remote control, RPA or recorder is documented in the docs, the integration pages or the release notes.

Sourcedocs.serval.com RSA SecurID and Presidio integration pages, with the release notes indexread 2026-09-12