AI harness software is what turns a language model into a working agent. Devin is one such harness from Cognition: an AI software engineer available as a cloud-based agent with its own editor, terminal, and browser, and as a local Devin CLI for quick terminal tasks. As of 20 September 2026.
What is Devin?
Cognition introduced Devin in March 2024 as “the first AI software engineer” – a, by its own description, fully autonomous AI agent that handles tickets, implements features, fixes bugs, and builds tools. Access runs through the web app at app.devin.ai, complemented by integrations with Slack, Microsoft Teams, and the Devin API. In April 2026, Cognition added a local variant with the Devin CLI: it installs on macOS, Linux, and via WSL with curl -fsSL https://cli.devin.ai/install.sh | bash (macOS also via Homebrew, Windows via a PowerShell script) and runs directly in the terminal for quick fixes, code exploration, and interactive coding.
Devin as an AI harness
Cloud Devin distinguishes an Ask mode for planning together and an Agent mode for autonomous execution; repositories must be indexed before the first session so Devin understands the codebase. Project context comes from an AGENTS.md file with setup commands, code standards, and test requirements, which Devin automatically looks for before starting work. The Devin CLI adds its own permission model with Allow/Ask/Deny rules and a sandbox that, according to the documentation, refuses to start if sandbox resolution fails rather than running unsandboxed. For scoped subtasks, the CLI lets you define custom subagents as Markdown files under agents/, running in the foreground or background. The official Devin MCP server (https://mcp.devin.ai/mcp) lets other MCP-capable tools manage sessions, playbooks, knowledge, and schedules; conversely, the CLI itself can use external MCP servers as tools.
Models and model access
According to the documentation, the Devin CLI supports models from Anthropic, OpenAI, Google, and Cognition itself, as well as open-source models such as DeepSeek, Kimi, and GLM; short names such as opus, sonnet, swe, codex, or gemini each resolve to the current version of that model family, including Claude models and GPT models. Users switch models via the –model flag, the /model command, or the configuration file; per Cognition’s own enterprise FAQ, Devin does not currently support custom LLM API keys or model endpoints, as it is built as a closed “compound AI system”. Devin’s model access therefore cannot be routed through a central proxy such as the AI Gateway.
Enterprise use: data protection and governance
For Enterprise customers, alongside the shared cloud, Cognition offers a dedicated deployment in a single-tenant VPC connected via AWS PrivateLink or an IPSec tunnel; according to the documentation, customer data stays within the customer’s tenant and is encrypted in transit and at rest. Enterprise organizations connect identity via SAML or OIDC (including Okta and Microsoft Entra ID) and sync users via SCIM 2.0. A genuine on-premises installation without Cognition’s infrastructure is not documented. Important for enterprises: source code and prompts go to Devin or the selected model provider and should be reflected in the records of processing activities beforehand.
Orchestrating Devin with herdr
herdr explicitly lists the Devin CLI (not the cloud agent) in its agent table with the integration level “screen manifest” and “session”. This means herdr detects local Devin CLI sessions via the terminal screen manifest and folds them into its session management; the cloud-based Devin agent itself runs outside herdr’s terminal context.
Where it fits: developer harness and enterprise harness
Devin targets development teams who want to delegate tasks to an autonomous agent or need quick fixes in the terminal. For non-developer staff, CompanyGPT is the matching harness: the same underlying idea of tools, context, and permissions, but built for enterprise roles, integrations, and audit requirements. The AI Gateway can serve as a shared cost and governance layer across different harnesses. For an overview of other coding agents and orchestrators, see our overview of all AI harnesses. We help companies introduce coding harnesses like Devin safely and connect them to their existing AI governance.