Backplane: a self-hosted project board shared by people and coding agents

Backplane gives a team and its connected AI tools a shared project record. Tasks, decisions and review results stay on the same board so a new contributor or restarted agent session can pick up the work with its context. The open-source preview includes a self-hosted platform, an MCP server and experimental runners.

Start with a board your team can operate

The Backplane product page describes Kanban cards with owners, priorities and dependencies, alongside notes, files and project definitions. The platform works without AI connections. A team can use the board to decide what needs doing and inspect what reached review before adding automated execution.

The current setup starts from the repository and Docker Compose. The publisher says the Compose deployment starts PostgreSQL, the backend and frontend, with required local configuration before creating an administrator account. Its self-hosting guide covers remote access and securing a public instance. Managed Cloud and Enterprise offerings are marked as planned; they should not be treated as available hosted subscriptions.

Give connected tools a scoped view

Backplane's included MCP server connects compatible clients to the project workspace. The page describes scoped toolsets for an interactive client, a role or a runner stage, rather than loading the entire tool catalog into every session. The server enforces the tools exposed to that session.

That division lets a team choose what a reviewer or implementer can access while keeping the underlying board shared. Review the actual permissions and connection configuration before giving an agent access to private project files or production-related work. Scoping a toolset is one control, not evidence that every subsequent action will be appropriate.

Add runners with explicit review and limits

Backplane says its experimental runners execute Claude Code or Codex CLI through the team's own accounts and rules. The connected Claude Code coding agent can read project code, edit files and run commands; Backplane supplies the shared board, role configuration and review history around that work. It does not include or resell an AI model.

Runner roles can define who plans, implements and reviews, with checks for tests, merge conflicts and overlapping changes. Experimental execution loops add a total spending limit, an attempt limit and a cutoff after repeated failures. The page says loops reread configuration at each iteration. Those controls need a concrete task and acceptance criteria; a completed card alone does not establish that its code or result is correct.

The publisher's project counts and autonomous-run examples describe its own use, not an independently reproduced success rate. For a first evaluation, use a small repository task, require a human to inspect the diff and tests, and check whether the board retains enough evidence for the next person to understand the result. Keep ownership and approval rules visible before expanding the runner's scope.