banksia

Install and start

Banksia requires Python 3.12 or newer and currently supports Linux and macOS. Install the command-line application in an isolated environment with pipx:

pipx install banksia
banksia init
banksia serve

Open http://127.0.0.1:18125/. The installed package includes the visual Console. Guided initialization chooses a default workspace, configures a Task provider, and can configure the separate Operator that helps you build and run teams.

Banksia uses SQLite by default, so a local installation needs no database server. For PostgreSQL, install the optional driver and supply a SQLAlchemy URL during initialization:

pipx install "banksia[postgres]"
banksia init \
  --database-url "postgresql+asyncpg://banksia@127.0.0.1/banksia"

See Getting started for provider prerequisites and a complete first run, or Database configuration for PostgreSQL permissions, schemas, and environment overrides.

Design and operate visually

The Console is the primary Banksia experience. It keeps the common path no-code while preserving advanced provider, sandbox, Skill/MCP, and capability controls when a team needs them.

• Workflow library makes reusable teams and drafts easy to find, compare, start, or remove.
• Workflow Studio shows the complete responsibility hierarchy on a horizontal canvas. Add a child with one +, select any Member to edit its purpose and instructions, validate the draft, and publish an immutable revision explicitly.
• Run Studio shows the live team, current plan, meaningful Activity, requests that need your attention, managed Actions, referenced files, and the exact completed or blocked Result.

The Operator is a separate conversational agent inside the Console. Ask it to draft or revise a Workflow, explain the current team, validate and publish when you request it, start and control Runs, answer Human Requests, or inspect managed Action output. It uses the same controller-owned operations as the visual interface, so chat does not create a second hidden source of truth.

Why Banksia

Banksia makes complex multi-agent work easier to create, adapt, recover, and trust:

• Create an AI team easily. Use the visual Console or conversational Operator to shape, publish, and run a reusable team.
• Let the team adapt how it works. Managers choose sequential, parallel, iterative, batch, or hybrid work from the Task and current evidence.
• Keep complex work moving. Durable state lets eligible work pause, wait, replan, retry safely, recover, and resume without discarding accepted history.
• Stay accountable from start to Result. Follow explicit responsibility, team revisions, decisions, Activity, referenced files, managed Actions, and the one Result accepted by the lead.

What a team can do

A Workflow defines reusable responsibility rather than a fixed schedule. You publish a team, give its lead one complete prompt, and let each Manager choose the approach that fits the work. If the current team itself no longer fits, a bounded replan changes its responsibility tree while preserving earlier revisions and completed work.

• Delegate through a responsibility tree without forcing a fixed sequence into the Workflow.
• Combine independent research, implementation, criticism, integration, and verification.
• Ask you a typed question when an authorized Member genuinely needs a decision.
• Run a managed command when the Workflow explicitly grants that capability.
• Replan a current responsibility subtree without rewriting the Task's earlier team or work history.
• Preserve progress across browser or provider interruption.
• Return one human-readable Result with direct references to detailed workspace files.

Capabilities deny by default and never inherit from a parent. Provider, model, sandbox, and capability choices remain available when a team needs them, without making them mandatory for every Workflow.

Upgrade ad-hoc delegation

If you already coordinate subagents by hand, Banksia turns those recurring delegation patterns into reusable teams without making subagents the product boundary:

Start with a proven team

banksia init publishes eight provider-neutral Starters. Choose one when the work is consequential or broad enough that independent responsibility can improve the outcome:

The Starter catalog includes example missions, expected deliverables, and guidance on when a simpler team is better.

How a run works

1. Choose or design a team. Start from a Starter or shape a responsibility tree in Workflow Studio.
2. Publish a stable revision. Every run keeps the exact team contract it started with.
3. Give the lead one complete prompt. The team plans, delegates, adapts, and records meaningful progress while detailed work stays in ordinary workspace files.
4. Follow the work and receive the Result. Answer real Human Requests, inspect managed Actions, and read the lead's exact final outcome.

Know before you start

Banksia is best suited to complex developer and researcher work on a trusted local machine. The controller runs as one loopback-bound process, and every Member in a Task shares one provider-visible workspace.

Codex and Claude are managed providers. Trusted full-access Task Members can use enabled user and project Skills plus configured MCP servers from those providers; narrower execution is isolated automatically, and Operator always stays isolated. Banksia does not yet define, install, or manage Skills, MCP servers, or general plugins itself.

OpenClaw is a user-operated compatibility transport. Native Windows is not currently supported; WSL2 uses the Linux path. Distributed delivery, broad multi-user operation, and per-Member isolated workspaces are outside the current product boundary.

Documentation

• Getting started
• Understand Workflows and teams
• [Author a Workflow](docs/gu