LinearB

Agent actions land in GitHub, including Enterprise Server, in GitLab SaaS and self managed, and in Bitbucket Cloud: gitStream approves, merges and closes pull requests, adds labels and reviewers, sets required approvals, requests changes, rewrites titles and descriptions, adds checks and runs GitHub workflows, all on write scope to the code. WorkerB approves small pull requests and opens tickets from Slack or Teams on a user's own Git login, and an HTTP request action reaches any other system. Labels are unavailable on Bitbucket and required approvals on GitLab.

Sourcedocs.gitstream.cm/automation-actionsread 2026-11-08

gitStream is the workflow engine. Rules written in YAML with Jinja expressions, kept in each repository or in one organization wide repository, match conditions on a pull request and run a list of actions in sequence, halting when an action fails. The AI review is itself an action in those rules, so its results feed the steps that follow it, which label the change, request reviewers or approve it. The engine runs inside the customer's CI or on LinearB's runner, and work is not split across several agents.

Sourcedocs.gitstream.cm/how-it-worksread 2026-11-08

Delivery data from Git, Jira, Azure Boards, incident tools and AI coding tools is normalized into a persistent store that merges each contributor's identities and links every pull request to its ticket, release and incident, kept for three years on Essentials and set against benchmarks from more than 8 million pull requests. Coding agents query that store through the MCP server for metrics, pull requests, branches, releases, incidents, services, teams and issues. Source code itself is never indexed or stored, so reviews read the diff and the team's review guidelines.

Sourcelinearb.io/use-case/software-factory-context-layerread 2026-11-08

Approval on a clean review is a setting that ships switched off, AI services as a whole sit behind an organization toggle only a company admin can change, and an issues limit caps how many findings one review posts. gitStream rules can require named reviewers, raise the number of approvals a pull request needs, or demand a human sign off on commits from coding agents it detects. Dry run mode and the gitStream playground show what a rule would do before it acts, and suggested fixes wait for a developer to commit them.

Sourcelinearb.helpdocs.io/article/9mpi9j3vk6-enabling-linearb-ai-servicesread 2026-11-08

LinearB holds SOC 1 Type 2, SOC 2 Type 2 and ISO/IEC 27001:2022, listed in its trust center beside GDPR, with Drata monitoring its controls. Sign in runs through SAML single sign on with Okta, Entra, JumpCloud, OneLogin or Google across up to 15 domains, with SCIM provisioning and MFA. Admin, Editor, Viewer and Basic roles are scoped by team, MCP access over OAuth follows each user's role, and MCP API keys are for admins only. Customer data trains no models and is not shared with model providers.

Sourcetrust.linearb.ioread 2026-11-08

An admin audit log exports as CSV or JSON, or through an export endpoint in the public API, and records every MCP tool call with the user, the tool, the argument names, the sign in method, the item count, the outcome and the client. Audit history is kept for one year on Essentials and for a custom period on Enterprise. Automation pages chart actions over time and pull requests processed by each rule, and the AI review dashboard tracks coverage, issues by category, issues resolved and suggested lines committed.

Sourcelinearb.helpdocs.io/article/ygz41u28cy-using-audit-logread 2026-11-08

The Agentic Studio assistant keeps a panel of past chats, and the AI reviewer carries its state across pushes on one pull request, withdrawing its own comments once a later commit makes them outdated. Review guidelines in REVIEW_RULES.md files are rules people write rather than anything the AI reviewer retains. No memory store holds what an agent learned for a user, team or repository with a set lifetime or a way to clear it, and thumbs up and down in the assistant improve the product rather than a team's own context.

Sourcelinearb.helpdocs.io/article/z9niuhn19i-agentic-studio-assistant-homeread 2026-11-08

The service runs in LinearB's cloud on AWS with no choice of region or data residency. Two parts can sit closer to the customer. An on premises agent for Enterprise runs on Kubernetes inside the customer's network, collecting data and sending it on to the LinearB cloud, and gitStream in repository mode runs its rules in the customer's own GitHub Actions or CI, while AI review still calls LinearB's service. Managed mode runs on LinearB's runner for GitHub, and LinearB Ltd. in Tel Aviv processes data as an affiliate.

Sourcedocs.linearb.io/oparead 2026-11-08

An automation library and an in app marketplace offer ready made gitStream rules to adopt as they are or edit. Seven managed automations switch on without writing a rule: AI review, AI descriptions, labels that flag code written by agents, estimated review time, missing Jira links, and automatic approval of Dependabot minor and patch updates. A prompt library holds ready made MCP prompts, and catalog dashboards cover AI return on investment, engineering health and the software factory.

Sourcedocs.gitstream.cm/automations/automation-libraryread 2026-11-08

Pull request events start the work: a new pull request, each commit, comments, title and description edits, labels added or removed, readiness for review, approval, merge, close and reopen. On GitHub, developers also call /gs review, /gs desc and /gs experts in a comment, and WorkerB responds in Slack and Teams. Automatic AI review runs on the first three pushes and skips drafts, bot authors and changes above 3,200 lines. Nothing inside LinearB starts agent work on a schedule; scheduled MCP prompts run in the customer's own client.

Sourcedocs.gitstream.cm/execution-modelread 2026-11-08

AI review, pull request descriptions and the assistant run on Anthropic's Claude Sonnet 4.5 through AWS Bedrock, and no setting lets a customer pick another model, route tasks between models or bring its own key. The homepage shows LinearB suggesting a cheaper model and effort level for a coding agent's task, yet none of the MCP server's 25 tools makes such a recommendation. Model choice for the customer's own coding agents stays inside those tools.

Sourcelinearb.helpdocs.io/article/uyzvby9nu1-supported-languages-in-ai-code-reviewread 2026-11-08

A public REST API takes an API key in a request header, allows 60 calls a minute on each endpoint, and covers measurements, custom metrics, deployments, incidents, services, repositories, teams, users with create, update and delete, API tokens and the audit log, so outside systems can report deployments and manage the organization. A remote MCP server signs in through OAuth or an admin key and exposes 25 tools, every one of them a read. No client libraries ship.

Sourcedocs.linearb.io/api-overviewread 2026-11-08

The AI Playground, in closed beta for admins, runs review and description guidelines against real pull requests before they are saved. Usage metrics show how many AI findings developers resolved and how many suggested lines they committed, and AI analytics compare assisted and human pull requests on cycle time and cost per effective pull request, measuring outcomes rather than running scored test cases. gitStream dry runs show what a changed rule would do to a pull request.

Sourcelinearb.helpdocs.io/article/tuyalymj6j-ai-playgroundread 2026-11-08

LinearB works through Git provider APIs, Slack and Teams bots, a REST API and an MCP server. None of its agents opens a browser, drives a desktop or clicks through a web interface: the AI reviewer reads diffs, gitStream acts through provider APIs, and the assistant queries LinearB's own data. When a customer's coding agents use a browser, that work stays inside those tools and reaches LinearB only as usage data.

Sourcelinearb.io/platform/workflow-automationsread 2026-11-08