Cognition

Named integrations include GitHub, Linear, Slack, Microsoft Teams, Datadog, Sentry, AWS, Azure, Snowflake, Databricks, MongoDB, PostgreSQL, Notion, Asana, Airtable, Segment, Stripe and Google Drive, with the vendor stating Devin works with hundreds of tools. GitHub integration ships pull requests the way the team does, picking up review feedback and CI results to get each one approved and merged; tickets are assigned in Linear directly or by label; Devin is tagged in Slack or Teams conversations to surface context or turn a discussion into a pull request.

Sourcedevin.airead 2026-08-30

Devin spins up a team of Devins for large tasks and assigns a fleet of agents to migrate all repositories in parallel, supporting multi-week, multi-repository projects with complete auditability at each step; a single task runs plan, code, test and pull request end to end in a cloud sandbox, and Devin Desktop's Agent Command Center presents local and cloud agents in one view of what is running, awaiting review and done, with Spaces letting multiple agents share context and Git worktrees.

Sourcedevin.ai and cognition.com/blog/introducing-devin-desktopread 2026-08-30

Devin learns the customer's codebase and picks up tribal knowledge, with new knowledge surfaced for human approval before it is retained and scoped to context such as working in a particular repository; DeepWiki auto-generates documentation and system diagrams for legacy codebases giving visibility into systems the team did not build, and is published publicly at deepwiki.com. Devin Desktop adds Codemaps for visualising code structure and Fast Context for retrieving the exact files an agent needs, and Devin is built for complex multi-repository projects.

Sourcedevin.ai, deepwiki.com and cognition.com/blog/introducing-devin-desktopread 2026-08-30

Devin Review ships configurable run controls of manual only, run when a pull request is first opened, or run on every commit, with the vendor stating customers should decide how often automatic work happens and migrating existing every-commit users to the pull-request-open default; new knowledge Devin picks up is surfaced for human approval before retention and is scoped to a stated context; work is delivered as a pull request for review, and Devin Desktop's Agent Command Center presents a single view of what is running, awaiting review and done.

Sourcedevin.ai and cognition.com/blog/new-self-serve-plans-for-devinread 2026-08-30

A published Trust Center documents security practices, compliance certifications and data privacy commitments, with SOC 2 Type 2 stated by the vendor, FedRAMP High reported as in-process from July 2026 and a dedicated federal compliance page.

Named controls documented individually: single sign-on via Okta, Microsoft Entra ID, generic SAML 2.0 and generic OIDC; SCIM 2.0 user and group provisioning; IdP group integration; custom roles and role-based access control; IP access lists restricting API and web app access by CIDR; customer-managed AWS KMS keys for encryption at rest on Enterprise Assured; secrets and site cookies management; OIDC short-lived keyless cloud access; security profiles restricting network, MCP, git and GitHub CLI access bindable to orgs, automations and sessions; AI guardrails screening messages for prompt injection, data exfiltration and policy violations; and attribution filtering screening generated code against public repositories and blocking matches before they are written. Data is encrypted in transit and at rest with each session on an isolated machine. Trust centre contents were not retrieved.

Sourcedocs.devin.ai/enterprise/security-access/trust-center, docs.devin.ai/enterprise/features/ai-guardrails, docs.devin.ai/enterprise/features/attribution-filtering and docs.devin.ai/product-guides/security-profilesread 2026-08-30

Session Insights analyses a customer's Devin sessions and returns actionable feedback to improve future interactions, so sessions persist as inspectable objects; ACU consumption is metered per session with a documented usage guide, enterprise billing analytics tracking consumption and per-user limits, Personal Analytics letting users view their own consumption, and a federal service-key API exposing team, group and per-user ACU consumption over historical windows.

Devin Coach adds analytics on suggestion outcomes and blockers. Migration work is marketed as carrying complete auditability at each step, and per-session audit logs are reported by a Cognition enablement partner.

Sourcedocs.devin.ai/product-guides/session-insights, docs.devin.ai/admin/billing/usage and docs.devin.ai/enterprise/security-access/personal-analyticsread 2026-08-30

The vendor states Devin gets better over time by reading past session trajectories, so prior runs are retained and inform later ones; a curated knowledge layer persists tribal knowledge approved by the team and scoped to context such as a particular repository, and Devin Desktop's Spaces let multiple agents share context and Git worktrees across related agents and cloud sessions.

Sourcedevin.ai and cognition.com/blog/introducing-devin-desktopread 2026-08-30

Devin Outposts runs sessions on the customer's own infrastructure with self-hosted workers, documented across an overview, quickstart, orchestration guide, API and CLI reference and partner integrations.

Two hosted models are documented alongside it: Enterprise Cloud, where brain and Devbox run in Cognition's multi-tenant cloud with each session on its own isolated machine, and Customer Dedicated Deployment, where the Devbox runs in a customer-isolated single-tenant VPC connected to the customer's VPC by AWS PrivateLink or IPSec with MFA VPN and OpenVPN support; Enterprise Assured adds customer-managed AWS KMS keys for encryption at rest.

Separate federal deployments exist, and GitHub Enterprise Cloud with Data Residency is supported. In the hosted models the brain always resides in Cognition's cloud as a stateless service, requiring egress over HTTPS and a persistent secure WebSocket.

Sourcedocs.devin.ai/cloud/outposts/overview, docs.devin.ai/enterprise/deployment/overview and docs.devin.ai/enterprise/security-access/security/customer-managed-keysread 2026-08-30

A Plugin marketplace lets administrators install and require bundles of skills for everyone in an organisation or enterprise from the Devin web app, with a documented plugin ecosystem for building, hosting and governing a shared set under policy controls, and a team marketplace quickstart for shared skills, rules, hooks and MCP servers. Playbooks are documented as building a library of reusable prompts for the organisation, with separate creating and using guides; Devin Skills teach reusable procedures through committed SKILL.md files; an MCP Marketplace is published, alongside a use-case gallery, a tutorial library and a prompt templates cheat sheet.

Sourcedocs.devin.ai/product-guides/plugins, docs.devin.ai/product-guides/creating-playbooks, docs.devin.ai/product-guides/skills and docs.devin.ai/work-with-devin/mcpread 2026-08-30

Work reaches Devin from Datadog incidents, Slack and Teams mentions, Linear tickets or labels, GitHub events and CI failures, and can be scheduled for recurring chores such as daily QA, release notes and documentation maintenance; the Devin API and Devin Automations allow full programmatic invocation. Devin Review supports configurable run triggers of manual only, run when a pull request is first opened, or run on every commit, with the vendor migrating existing every-commit users to pull-request-open by default.

Sourcedevin.ai and cognition.com/blog/new-self-serve-plans-for-devinread 2026-08-30

Devin documents a dedicated Models page covering available models and how to configure them, an Adaptive model router described as Cognition's intelligent router that automatically selects the best AI model for each task, and federal Model Provisioning controlling Cascade, Command and CLI model availability by team and group, with group management exposing model availability alongside ACU limits.

Cognition also ships its own SWE series models. Bring-your-own-key is explicitly refused: the enterprise deployment FAQ states Devin is a compound AI system and does not currently support third-party LLM API keys. The specific provider list was not retrieved.

Sourcedocs.devin.ai/cli/models, docs.devin.ai/cli/adaptive, docs.devin.ai/federal/model-provisioning and docs.devin.ai/enterprise/deployment/overviewread 2026-08-30

A public Devin API allows full programmatic control and is offered alongside Devin Automations for automating work without code, with the vendor positioning both as the path to automating entirely; Devin Desktop supports the Agent Client Protocol so agents interoperate across clients, and Cognition provides an SDK for developers building on Devin. API access is documented as a Teams and Enterprise tier capability rather than available on lower self-serve plans.

Sourcedevin.ai and cognition.com/blog/introducing-devin-desktopread 2026-08-30

Devin performs unit and end-to-end testing, visual QA with browser and desktop use on its own output, automatic CI failure fixing, and performance optimisation, and Devin Review identifies and resolves bugs in pull requests; a newly launched Security Swarm product reviews code for security issues. These verify the customer's code and the agent's own output rather than providing an evaluation harness the customer points at agent behaviour, and no agent benchmarking or regression surface is exposed to customers.

Sourcedevin.ai and devin.ai/securityread 2026-08-30

Visual QA is listed as a headline use case with full browser and desktop use, alongside repetitive browser task automation, web research and scraping, and unit and end-to-end testing; Devin's sandbox is described as including a shell, a browser and an editor, and browser actions are named in the vendor's own definition of the Agent Compute Unit as a billable component of a run.

Sourcedevin.ai and cognition.com/blog/new-self-serve-plans-for-devinread 2026-08-30