Outreach

Deep in one direction and open in another, which is the shape of a platform built to sit on top of a CRM. The architectural assumption is Salesforce: account and opportunity data pulled in, activity pushed back, with the platform designed to complement rather than replace the system of record. Around it sit mailbox and calendar connectivity for sequence execution, a dialer for calling, and Kaia for meeting capture.

What lifts this beyond a fixed connector list is the marketplace: third-party apps and MCP connectors are built against a published spec, reviewed by Outreach for protocol compliance, scoping, performance and security, and installed by admins from the Outreach Marketplace, so the integration surface extends without the vendor building each one. The REST API covers tasks, templates, call logging, mailboxes, webhooks, bulk actions and custom objects, and custom objects mean a customer can model their own entities into the platform.

Sourcedevelopers.outreach.io/api/reference/overview and support.outreach.io MCP Connectorsread 2026-09-04

Named agents running multi-step work over an established sequence engine, with the chain visible in the vendor's own tool documentation. Revenue Agent identifies accounts, sources contacts and generates outreach end to end; Research Agent pulls account research from internal and external sources on a set cadence and writes findings into Outreach account fields; Deal Agent derives CRM updates from conversation intelligence without rep intervention; a Meeting Prep Agent and a Reply Agent extend the set.

Teams can customize the workflows to mirror how their high performers work rather than accepting a fixed path. The developer documentation shows the same orchestration exposed outward, with published tool chains running prospect creation into sequence enrollment into sequence-state readback, and rulesets attached to sequences governing how they execute. The chain assumes prospect data already exists, because Outreach carries no contact or company database of its own.

Sourcedevelopers.outreach.io/mcp-server/usage-guide and the December 2024 AI Prospecting Agents announcementread 2026-09-04

Research assembled per account rather than retrieval over a corpus the customer owns. Research Agent draws on first-party and third-party sources, including CRM detail and recent news, and writes what it finds into structured Outreach account fields, and the MCP layer exposes an account question-answering tool over the same account context. That is grounding on the customer's records plus external research, and it is durable because the output lands in fields rather than a chat window.

It is aggregation into a schema rather than a documented retrieval layer over the customer's own content: nothing lets a team connect a document repository, wiki, knowledge base or content library, no citation or source attribution is documented on generated research, and the platform carries no contact or company database of its own, so the external half depends on data the customer brings from elsewhere.

Sourcedevelopers.outreach.io/api/reference/overview and the December 2024 AI Prospecting Agents announcementread 2026-09-04

A real autonomy control and configurable governance, with the platform explicitly declining to provide the gate.

What the customer holds: a first-class choice between running a capability as a copilot that assists a rep and running it as a fully autonomous agent, stated by the vendor at launch, so the autonomy level is a deployment decision rather than a fixed property; and rulesets, which Outreach describes as governance rules attached to a sequence and exposes as a searchable object, so policy constraints on what a sequence may do are configured per organization.

Access to agent capability is further bounded by license, with the vendor enforcing that a user must be signed in, active and licensed before any data moves.

No approval queue, review step, confidence threshold or escalation rule is documented, and on the extensibility path the vendor says the opposite outright, instructing developers to build confirmation flows, audit logging and rate limits in their own agent layer because the MCP server does not throttle or undo, and to use destructive action hints to drive their own confirmation experience. That is the platform telling integrators the gate is theirs to build.

Sourcesupport.outreach.io MCP Server Overview, developers.outreach.io/mcp-server/usage-guide and the December 2024 AI Prospecting Agents announcementread 2026-09-04

The compliance record and the access model are both documented. Outreach holds SOC 2 Type II, ISO 27001 and ISO 27701, TRUSTe certification, Microsoft 365 App Certification and a CSA STAR entry, with DPAs carrying EU and UK standard contractual clauses, an OWASP-aligned lifecycle attested in the SOC 2 report and annual independent penetration testing. Identity is enforced at the agent boundary as well as the console: no data moves over the MCP tools unless the user is signed in, holds an active account and a valid license, and access is scoped per connecting identity.

Sourceoutreach.io/platform/security and support.outreach.io MCP Server Overviewread 2026-10-01

Managers can see the pipeline; nobody can see what the agents did. Deal health scoring, forecasting and Kaia conversation intelligence give visibility into pipeline movement and rep activity, and sequence performance reporting shows what engagement produced.

None of that reaches agent behavior, and the vendor's own developer documentation makes the gap explicit: it instructs integrators to build audit logging into their own agent layer, and states that the MCP server does not throttle or undo. A platform that tells developers to log agent actions themselves is a platform not logging them centrally.

Nothing documents a trace of what Revenue Agent prospected and why, what Research Agent read before writing a finding into an account field, or which records Deal Agent updated from a call, and no configuration change trail or retention control appears.

Sourcedevelopers.outreach.io/mcp-server/usage-guide and outreach.io platform pagesread 2026-09-04

Agent output written into records, not memory. Research Agent writes its findings into Outreach account fields and Deal Agent updates the CRM, which makes them durable business records; no memory the agents keep and read back as their own state, with a scope, lifetime or delete path, is documented.

Sourcedevelopers.outreach.io and support.outreach.io Agents & AI Overviewread 2026-10-01

Multi-tenant cloud with no location or topology control offered. No region selection, EU or other in region option, private cloud, VPC, single tenant tier or on premises route is documented. What the vendor documents on cross-border movement is legal cover: data processing agreements carrying EU and UK standard contractual clauses executed with vendors processing customer data, and a Privacy Shield reference. Legal cover is not residency. This is an enterprise platform holding ISO 27701 and selling into organizations that routinely ask for regional hosting, so the absence is conspicuous.

Sourceoutreach.io/platform/securityread 2026-09-04

Named agents sold by package plus a template set. The pricing page places Omni, Research, Deal and Meeting Prep agents in Amplify Essentials and adds the Revenue and Personalization agents in Core, each with its own job, and Agent Studio offers eleven pre-curated workflow templates a customer selects and configures.

Sourceoutreach.ai/pricing and support.outreach.io Agent Studio Overviewread 2026-10-01

Agents act without a person prompting them. Agent Studio workflows start on events such as a new prospect or an updated opportunity, or at scheduled times, and the Research Agent runs research on a set cadence; work goes out across email, calls and LinkedIn content.

Sourcesupport.outreach.io Agent Studio Overview (updated 12 May 2025) and Agents & AI Overviewread 2026-10-01

The customer does not choose the model and no provider is named. No model list, routing policy, per agent selection or bring your own key option is documented, and the AI layer is presented as trained on Outreach's own engagement signal corpus, which is a claim about training data rather than about which model runs. The MCP server lets a customer's own assistant, Claude among the named clients, call Outreach and act through 32 tools, but that is the customer picking the interface that reads the vendor, the opposite of choosing what powers the product.

Sourceoutreach.io/platform and developers.outreach.ioread 2026-09-04

A documented API and an MCP server for the vendor's own platform. developers.outreach.io publishes a REST API reference (OAuth 2.0 and a server to server flow; tasks, templates, calls, mailboxes, webhooks, bulk actions, custom objects), with API call allowances by package, and a hosted MCP server at api.outreach.io/mcp with 32 tools including creation and sequence enrollment under OAuth 2.1.

Sourcedevelopers.outreach.io and outreach.ai/pricingread 2026-10-01

An optimization loop over outbound performance, aimed at the outcome rather than at the agent. Sequence performance analytics measure what engagement produced, deal health scoring assesses opportunities against criteria the team can read, and the AI layer is described as learning from engagement outcomes across the platform's own signal corpus, so what is measured feeds back into what is recommended.

The criterion is the business outcome. Nothing is aimed at the agents themselves: no evaluation harness, scored test set, benchmark, simulation, dry run or regression check when an agent's configuration changes is documented, so a change to how Revenue Agent prospects is judged only after it has reached real buyers.

Sourceoutreach.io/platformread 2026-09-04

Everything the agents touch arrives through an authenticated interface, and the vendor's own tool inventory demonstrates it.

Work reaches the CRM through the Salesforce integration, email through connected mailboxes, calls through the dialer, meetings through calendar connections, and external systems through the REST API, webhooks and the MCP server, whose 32 published tools are all record operations rather than interface operations: search, retrieval, creation, deletion and sequence enrollment.

No browser extension driving third-party software, headless session, portal navigation, form filling, screen or click automation is documented. The developer reference enumerates exactly what the platform can do from outside, and none of it is interface control. Nothing like an extension that fills questionnaires inside third party portals is documented either.

Sourcedevelopers.outreach.io/api/reference/overview and outreach.io/platformread 2026-09-04