Features

What OutcomeOps Does

Grounded PR review and an organizational knowledge base, with optional code generation. Open source under Apache 2.0, deployed in your own AWS account.

PR Review and Governance

Grounded pull request review judged against your own ADRs and standards, with an audit trail. The most-adopted capability, and the place most teams start.

Each PR gets ADR compliance, breaking changes, architectural duplication, README freshness, test coverage, and license/copyleft checks, each with a PASS, WARN, or FAIL verdict.

The ADR, duplication, and license checks retrieve the standards that apply to the change from your knowledge base and judge the diff against them. There are no hardcoded rules: change an ADR and the review changes with it.

Before the checks run, OutcomeOps posts a short prose walkthrough of the diff itself, not a restatement of the PR description. It is updated in place on every push instead of piling up new comments.

Findings are posted on the exact lines they refer to, alongside the per-check summary. Comments are updated when the code changes and resolved when the finding no longer applies.

For every symbol a PR changes, OutcomeOps looks up its callers and importers in the symbol graph as of the PR’s base commit and folds that impact into the review. A signature change to a shared library flags the consuming repositories, not just callers in the same repo.

Empty results are reported honestly: a symbol with no callers is shown as “no known callers” rather than omitted or guessed. If the graph is unavailable or stale, the review says so.

GitHub pull requests, GitLab merge requests, and Azure DevOps pull requests get the same checks and the same comments, upserted in place on each push. Reviews are triggered by a CI workflow you install (a GitHub Action using OIDC, or the GitLab and Azure DevOps equivalents) when a PR is opened or updated, including PRs written by AI.

Standards placed in a global workspace are applied to every check in every repository, so security and architecture rules defined once reach every team’s PRs.

Organizational Knowledge Base and Chat

Everything your teams have decided, documented, and built, searchable in one place, with answers that cite where they came from.

Content is embedded with Amazon Titan Text Embeddings v2, stored in Amazon S3 Vectors, and reranked with Cohere Rerank 3.5 on Bedrock. There is no third-party vector database to procure or operate.

Anthropic Claude and OpenAI GPT models on Amazon Bedrock. The model for each tier is a configuration setting, so you can run Sonnet and Haiku, move the advanced tier to Opus, or use GPT-5.5, without code changes.

  • File and line citations are resolved to the enclosing function or class, so they point at the code that matters.
  • The model can read the exact file at the exact commit from GitHub, GitLab, or Azure DevOps to verify a claim before making it.
  • Code and documentation are retrieved separately and ranked on their own, so stale docs cannot outvote what the code actually does.
  • List-style answers never invent totals: every item is enumerated from sources, or no count is claimed.

Structural questions such as “who calls this?” are answered from the symbol graph with exact file and line results, while open-ended questions use semantic search. Every answer shows which path was taken and how fresh the graph data is, so reviewers can tell deterministic graph results from search results.

A streaming chat interface with citations, file attachments, saved conversations, sharing with a workspace, and search across every workspace you can access. Workspace admins can set a custom system prompt for their workspace.

Source Integrations

Connect the systems where your organization’s knowledge already lives. Each source syncs on change and on a schedule.

  • GitHub via a GitHub App. GitHub Enterprise Server is configurable but not yet certified against a live instance.
  • GitLab, including gitlab.com and self-managed instances, several at once, with nested subgroups.
  • Azure DevOps Repos via Entra ID.

Jira issues and Azure DevOps Boards work items.

Confluence spaces, SharePoint sites, OneDrive, OneNote, Outlook email threads, Microsoft Teams channels, ServiceNow knowledge articles, and Salesforce accounts, contacts, opportunities, and activity.

Introspect PostgreSQL, MySQL, and SQL Server schemas as workspace context, with notifications when a schema drifts. Only schemas are read, never row data.

Drag and drop files into a workspace. Uploaded documents and attachments are run through Amazon Textract, so tables, scans, and image-heavy PDFs keep their content.

Code Understanding

Per-repository code maps and a symbol graph that turn your codebase into queryable architecture.

Java, Python, TypeScript (including React), Go, C#/.NET, and ABAP/SAP via abapGit.

Deterministic answers to “who calls this,” “who imports this package,” “which classes implement this interface,” and “where is this defined,” each with file and line. Bare names like require_auth resolve to every matching definition across the codebase.

When one repository uses a library another repository publishes, call sites link to the library’s real definition, across workspaces you have been granted access to. Supported for Python, TypeScript, Java, C#, and Go.

Plain library repositories are discovered with no configuration file. Code maps refresh incrementally from the commit delta, and an optional .outcomeops.yaml lets a repository tune what gets extracted.

MCP Server: Use It From Your Coding Agent

Bring the governed knowledge base to the AI tools your engineers already use, through the Model Context Protocol.

Claude Code, Cursor, and other MCP clients can query the knowledge base, read source at a commit, and run symbol-graph lookups, including a combined search that returns semantic and graph results together. Tool responses carry their citations, so the client model stays grounded.

Register additional MCP servers with OAuth 2.0 client credentials or JWT bearer authentication, and assign each one to the workspaces that should use it. Built-in Snyk and SonarQube servers add vulnerability and code-quality scanning.

Workspaces and Access Control

Organize knowledge by team, control who sees what, and sign in with your identity provider.

Each workspace has its own members, repositories, MCP servers, system prompt, monthly budget, and private vector partition. Workspaces can grant other workspaces search-only access.

Content in a global workspace appears in every user’s retrieval and every PR check, so org-wide standards are defined once.

Org admin, workspace admin, and member roles. The web UI signs in with Microsoft Entra ID (Azure AD) over OIDC; MCP and programmatic access use OAuth 2.0 tokens or API keys issued by the platform.

Audit, Retention, and Cost

Know who asked what, what it cost, and keep the record where your security team wants it.

Every model call and every admin and sign-in action is recorded with the user, timestamp, source IP, and user agent. The audit dashboard filters by workspace, user, and category, drills from a flagged response to that user’s history, and exports to CSV or JSON.

Optionally stream every audit event through Amazon Kinesis in OCSF 1.3.0 format to Splunk, Datadog, Sumo Logic, Amazon Security Lake, or your own consumer.

Chat history expires after 90 days by default and the audit dashboard keeps a 365-day window by default; both are configurable. Long-term retention lives in your SIEM through the export.

Spend by model, workspace, and user with daily trends, monthly budgets per workspace, and alerts on the day a threshold is crossed. Bedrock calls are tagged per workload so AWS Cost Explorer shows which workflow drives spend, and every price is configurable for negotiated AWS rates.

Security and Compliance

Controls for regulated environments, all running inside your AWS account.

Documents are scanned with Amazon Comprehend before they are embedded. Names, contact details, government IDs, financial and card numbers, passwords, and AWS keys are replaced with placeholders, so the model cannot reveal them. Each workspace can choose strict redaction or none, the organization can tune which types are redacted, and every ingested document records what was redacted.

Tokens and secrets are stored encrypted with AWS KMS, and data tables can use your own customer-managed key.

Secret scanning of connected repositories, plus Snyk vulnerability and SonarQube code-quality scanning through MCP.

Model refusals are captured in the audit log and alerted on. Users acknowledge an AI disclosure on first use, and the platform records who accepted which version and when. A data protection impact assessment (DPIA) document is included for GDPR and UK-DPA review.

Deployment and Operations

Terraform into your own AWS account, built on AWS managed services.

One deployment per AWS account, with no shared control plane and no connection back to any vendor. It is built on Lambda, DynamoDB, S3, S3 Vectors, SQS, SNS, EventBridge, KMS, and SSM, with only the web UI and MCP servers on Fargate. No EC2 and no Kubernetes.

The load balancer is internal by default and reached over VPC peering, Transit Gateway, Direct Connect, or VPN. A public option exists for trials and development.

Optionally deploy to a second region with application-level replication of every stateful resource, and fail over with a configuration flag, no Terraform apply required. See why OutcomeOps doesn’t use DynamoDB Global Tables.

Every Fargate service’s CPU and memory is configurable, Fargate Spot can be enabled for cost-sensitive environments, and a CLI supports ad-hoc knowledge base queries.

Code Generation (Optional)

An optional capability on the same knowledge base. Leave it off and use review plus the knowledge base, or point your own coding agent at the knowledge base through MCP instead.

A Jira ticket or a labelled GitHub issue becomes a plan, code, tests, and a pull request, grounded in your ADRs. A plan-only label produces just the plan for a few cents; full generation typically costs a few dollars per feature in Bedrock usage.

Python (AWS Lambda), Java (Spring), TypeScript (Express), and React. C#, Go, and ABAP are covered by code maps, the symbol graph, and PR review, but not by generation.

Generated code runs through its tests with automatic fixes for ADR, README, and test failures, and retries before giving up. If every attempt fails, you still get a pull request with the work so far. Every generated PR then goes through the same grounded review as human PRs.

This page summarizes each capability. The full, always-current catalogue lives with the code.