AnyForge is a platform for planning and delivering software with AI coding agents. Teams plan in programs, initiatives, specs, and work items. Crews of agents then build that work in the team’s own repositories, and a person approves each pull request before it merges. Around that loop sit an in-product assistant called the Oracle, a browser-based Code Studio, intake from Slack, and a Control API and MCP server so developers’ own coding agents can work against the same data.
I joined in June 2026 as a full stack engineer on a small team, and my work has touched most of the product since: the planning Console, the crews, the Oracle, Code Studio, the API and MCP surface, and the security and intake programs built on top of them.
What I built
- Planning and delivery in the Console: program pages with dependency maps across initiatives, initiative and team boards with lanes that follow label groups, team ticket codes, blocked work split into what is on us and what is waiting on others, cycles that can lock while still taking one-off scope changes, the Delivery queue, and report widgets for lead time and time in stage.
- The spec editor: rich text on Tiptap, drafts that survive a closed tab, Oracle-proposed edits, and whole-spec regeneration that runs as a server-held job and asks you to review the source evidence before it replaces anything.
- The crews, which run on LangGraph: repository conventions read from
AGENTS.mdandCLAUDE.md, structural code search, support for more languages, test-failure diagnostics, approved run profiles with reasoning effort set per role, and a benchmark mode that records each model’s billed cost. I also gave crews a Postgres test database and a disposable Docker host for Testcontainers suites, before and after screenshots on pull requests that change UI, pull requests that stay in draft until approval, and a way to ask a person to widen their write scope instead of halting. - Code Studio: the Build view with one file tree and one crew stream, read-only session sharing, the repository’s own skills and commands in the
/menu, and picking up a finished crew’s open pull request in a new session. - The Oracle: crew threads for running and following several crews from one conversation, file sharing and PDF exports, conversation search, a transcript-first meeting assistant, team notes any member can write, timers that post reminders and scheduled reports, and a terminal client in the
anyforgeCLI that talks to the same hosted engine. - The Control API and MCP server: parity with what Code Studio can do, an idempotency key on every write so a retried call never writes twice, connecting a REST API from its OpenAPI spec, hosted read-only servers for AWS operations and Figma and Penpot design files, private servers behind an owner policy, and approval before a crew calls a write tool. The server is published to the official MCP Registry.
- The security program: findings imported across repositories and branches, overlapping fixes folded into one group that one crew works, each finding routed to the team that owns its service, and a fix counted only once its pull request merges and, where the source supports it, the provider closes the finding.
- Intake and support: program-owned, versioned intake forms you can set up by talking to the Oracle, requests drafted from Slack threads, reminders that repeat until someone decides, a triage bot that assigns tickets to the least-loaded member of a team, and links from Datadog and PagerDuty pages back to the Slack thread and work item.
- Pull request review and skills: review lenses built from a company’s crew reviewers, a merge-risk rating for every pull request, a shadow mode that tries to disprove serious findings before trusting them, skill sources read at a pinned commit, and a scan for skill content that would weaken an agent’s safeguards.
Hard parts
Crews work unattended in other people’s repositories, so most of the effort goes into limits and checks around the agent. A crew gets a folder and a write scope, the repository’s CI toolchain caps which dependency fixes it may make, sibling crews see each other’s open pull requests, and every merge still needs a person. Each limit has a way out that a person can approve, such as widening the scope, lifting a base-merge hold, or accepting a named failing check, so a crew that hits one can ask for help.
The other hard part is that the same programs, initiatives, and work items are read and written from the Console, the Oracle, Slack, the CLI, and outside coding agents over MCP. Keeping those surfaces in step meant building API and MCP parity feature by feature, with idempotent writes and cursor paging, so an agent working through the API can do what a person does in the browser.
Stack
Next.js and React with TypeScript, Tailwind CSS and shadcn/ui, SWR, and Tiptap for rich text; Firebase with Firestore and Cloud Functions; LangGraph for the crews, with crew runners and the agent service on Google Cloud Run; Anthropic, OpenAI, and OpenRouter models; the Slack and GitHub APIs; PostHog; Playwright, Jest, and Storybook.