Banksia: Open-source platform for building multi-agent AI teams

Banksia is an open-source platform for designing and running multi-agent AI teams with full auditability and control. Installation is straightforward: Python 3.12 or newer, Linux or macOS, and a simple pipx install (or pipx install "banksia[postgres]" for PostgreSQL support). Running banksia init, banksia serve, and opening http://127.0.0.1:18125/ brings up the Console, a visual interface bundled with the package.

The core experience centers on two main interfaces. The Console is no-code for the common path but exposes advanced controls for provider, sandbox, and capability choices when needed. Within it, Workflow Studio displays the complete responsibility hierarchy on a horizontal canvas; users add child members with a single click, edit each member's purpose and instructions, validate, and publish immutable revisions. Run Studio shows the live team, current plan, meaningful activity, human requests needing attention, managed actions, referenced files, and the exact result. The Operator is a separate conversational agent that can draft or revise workflows, explain the current team, validate and publish, start and control runs, answer human requests, or inspect action output—using the same controller-owned operations as the visual interface, so chat does not create a hidden second source of truth.

A Workflow defines reusable responsibility rather than a fixed schedule. Managers can choose sequential, parallel, iterative, batch, or hybrid work based on the task and current evidence. Teams can pause, wait, replan, retry, recover, and resume without discarding accepted history, thanks to durable state. Capabilities deny by default and never inherit from a parent, so providers, models, sandboxes, and capabilities remain granular and explicit. The platform preserves progress across browser or provider interruptions and returns one human-readable result with direct references to detailed workspace files.

Banksia ships with eight provider-neutral starter templates: production-feature-delivery (for features crossing contracts and boundaries), incident-investigation-and-recovery (for competing hypotheses and verified fixes), migration-and-modernisation (for inventory and dependency-aware cutover), deep-research-and-decision-brief (for evidence-backed recommendations), decision-through-competing-prototypes (for fair testing of alternatives), idea-to-validated-demo (for evidence-backed product positioning), experiment-and-replication-program (for explicit methods and claim audit), and security-audit-and-hardening (for attack-surface mapping and adversarial re-verification). Each includes example missions, expected deliverables, and guidance on when a simpler team is better.

The tool is open source under the MIT License, except for the visual Console, which contains material derived from n8n and is distributed under the Sustainable Use License (permitting internal business, non-commercial, and personal use; redistribution limited to free, non-commercial distribution). Banksia is best suited to complex developer and researcher work on a trusted local machine; the controller runs as one loopback-bound process, every member in a task shares one provider-visible workspace, and external MCP servers, reusable skills, distributed delivery, broad multi-user operation, and per-member isolated workspaces are outside the current product boundary. Native Windows is not currently supported; WSL2 uses the Linux path.

Banksia addresses a real friction point in multi-agent AI workflows: the gap between ad-hoc coordination (manually recreating roles and prompts for each job, reconstructing ownership from transcripts, treating provider output as final) and structured, auditable team execution. The platform bridges this by letting teams be designed once and reused, with durable state that survives interruptions and supports replanning. Its explicit responsibility hierarchy and controller-owned activity log mean that every decision, handoff, and action is traceable—critical for high-stakes work where accountability matters.

The eight starter templates suggest the tool is aimed at consequential scenarios: feature delivery across boundaries, incident response, large migrations, security audits, and research-backed decisions. Each starter bundles a responsibility tree, expected deliverables, and guidance on when simpler teams are better, lowering the friction for teams to adopt structure without overengineering. By offering both a no-code visual Console and a conversational Operator, Banksia accommodates different user preferences while avoiding a hidden second channel of control.

No comments yet. Be the first to share your thoughts!