The platform and its Enterprise Edition are complete and merged. The long-term vision has four tiers; this is where each one stands.
v1.0Shipped
Repository Intelligence Platform
The committed near-term platform, delivered as three tracks and each closed against its own definition of done.
Core platform: 12-language parsing, Postgres, Neo4j, Qdrant and Redis or a single local file, retrieval, memory, workflows and agents
Compression & cost: the compression engine, Compress-Cache-Retrieve, the OpenAI/Anthropic proxy, savings tracking
Graph intelligence: confidence tags, the Neo4j graph engine, reflection and learning, full REST, MCP and CLI parity
The web console for people, served by the API
v2.0Shipped
Enterprise AI Platform
Running BBM-Atlas for many teams at once, as the Enterprise Edition: built one step at a time, each with its own tests.
Licence keys, signed and checked offline
People, service accounts and per-principal API keys, with five roles
Organisations kept apart on one deployment
Single sign-on with OIDC and SAML 2.0
An append-only, hash-chained audit log
High availability: replicas sharing one job queue, and a Helm profile
People, organisations, keys and the audit log managed from the web console
v3.0Planned
Distributed Engineering Intelligence
Reasoning that crosses repository boundaries.
Cross-repository reasoning
Distributed agent clusters
v4.0Vision
Engineering Operating System
The long-term ambition, with no committed scope yet.
Autonomous refactoring
Self-healing workflows
Up next
Before v3.0
BBM-Atlas 1.0.0 is on PyPI, npm, the VS Code Marketplace, Open VSX and GitHub's container registry: everything planned before v3.0 is out. The milestones above say what comes after.
Done
Recently closed
The latest work to land, newest first: the Enterprise Edition's last pieces, then the scale, operations and deferred scope that ran beside v1.0.
Released: the extension in Cursor and VSCodium
The VS Code extension's 1.0.0 on Open VSX, the registry Cursor, VSCodium, Devin Desktop, Gitpod and Theia install from, in BBM-Atlas's own verified namespace.
Released: BBM-Atlas 1.0.0
The first final release: the package on PyPI and npm, the VS Code extension's first stable version on the Marketplace, the container image and Helm chart, and the Claude Code plugin from this site's marketplace.
In 1.0.0: BBM-Atlas in Claude Code
A plugin brings BBM-Atlas's MCP server to every Claude Code session, with a skill that tells Claude which tool each task needs, and /bbm-atlas:index and /bbm-atlas:impact.
In 1.0.0: a re-index on save in about a second
After one file's edit, BBM-Atlas parses that file again, resolves only the links the edit can change, and writes only what changed: 7.7 s became about 1.3 s on its own 515 files.
In 1.0.0: Intel Macs
Every dependency installs from a prebuilt wheel on Intel Macs too, which CI checks with each change.
In 1.0.0: MCP in every client
bbm-atlas mcp install writes Cursor's, VS Code's and Gemini CLI's project files, and bbm-atlas mcp config prints the entry for Claude Desktop, Devin Desktop, Zed, JetBrains AI Assistant and Codex.
In 1.0.0: chat apps over OAuth
claude.ai, ChatGPT and the Gemini app can call BBM-Atlas's tools with a token from your authorization server - a JWT, or an opaque token asked about at the issuer - tested against a real one.
In 1.0.0: one token ledger for every replica
Replicas behind a load balancer record what BBM-Atlas saved and cost into one Postgres ledger, so the figures are the whole deployment's.
In the extension's 1.0.0: more in the editor
The graph beside your code, restored after a restart; paths between symbols, cycles and the most-connected symbols; facts; rated answers; token savings; and Report a Problem.
Released: the release candidate
BBM-Atlas 1.0.0rc1 on PyPI, one package for both editions, and the VS Code extension 0.3.0 on the Marketplace, as pre-releases.
Rate limits per key
Each API key, and each person signed in, has a budget of their own, so a team behind one office address doesn't share one.
Enterprise pages in the console
People, organisations, the audit log and your own keys, managed from the browser, each page shown only to the roles allowed to use it.
Single sign-on with SAML 2.0
For identity providers that speak SAML, alongside OIDC, with guides for Okta, Microsoft Entra ID, Google Workspace and Keycloak.
Performance at production scale
Impact analysis and the graph engine meet their latency targets on a 100k-node Neo4j graph, checked in CI.
Failure-log scanners per agent
bbm-atlas learn reads the session logs of Claude Code, Codex, Gemini CLI and OpenCode.
Production deployment
HTTPS through a Caddy overlay, credentials from files, a Helm chart, rate limits shared through Redis, and OpenTelemetry traces.
Deferred compression scope
Hybrid relevance scoring, an ML content detector and OpenTelemetry traces for compression.
Persistence with nothing to install
With no configuration, bbm-atlas serve keeps repositories, the graph and learned facts in one local file.
MCP over HTTP
The running API serves its tools at /mcp, so assistants share the repositories the console and the CLI see.
How it is built
Audited, tested, and tested again
Every change lands through a pull request with regression tests. CI runs the full suite against real Postgres, Neo4j, Qdrant and Redis, signs in through a real Keycloak, installs the Helm chart on a Kubernetes cluster, and smoke-tests the package on Linux, macOS and Windows.