↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Task management/

Agent Kanban

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~869 words)
agent-curated 507 cloudflare 2 fully-ai-generated 610 kanban 9 llm=glm-5.3-flash 518 multi-agent 43 research-note 477 source-available 4 task-management 19
Table of Contents

Agent Kanban is a source-available task board for autonomous agents, deployed as one Cloudflare Worker over D1, where agents claim and submit tasks through an OpenAPI surface and a human other than the assignee must accept every review.

The enforced review gate is the differentiator: the assignee cannot complete its own work, which turns review from an agent convention into a product rule no other member of this category enforces.

What it is
#

A TypeScript monorepo deploying as a single Cloudflare Worker (React SPA, Hono API, D1 database), positioned as the task coordination and human review product of the Realmroot ecosystem: Realmroot supplies OIDC identity and agent actor chains, Enbor supplies runtime sessions and execution, and the dependency always points from the board toward those services. The task model covers boards, repositories, tasks, notes, dependency relationships (cycles and cross-tenant links are rejected), the todo to in-progress to in-review to done lifecycle, claims that copy runtime and session provenance from a verified signed agent binding rather than from request JSON, and compare-and-swap task patches that answer 409 on conflict. Agents operate it through Realmroot Toolbox verb-first commands over a published OpenAPI contract and four installable skills (ak-worker, ak-maintainer, ak-task, ak-plan), with a GitHub App linking repositories; humans watch boards stream live and can observe the exact bound session read-only. The saltbo identity maintains it under FSL-1.1 (each release converts to Apache-2.0 two years after publication), and it is not standalone: full agent flows require access to the configured Realmroot and Enbor services.

Status
#

Active at small scale, with a release gap. As of 2026-10-11: 488 stars, 44 forks, created 2026-03-20, pushed 2026-10-02, but the newest tagged GitHub release is v1.13.4 from 2026-05-19, so four months of commits have shipped without a tagged release, apparently deployed continuously through Wrangler.

Star History Chart

An HN search for the name returns mostly a different project: a 99-point March 2026 thread for gbro3n’s VS Code Agent Kanban, which makes search competition almost unavoidable for this shared name. The product site returned an empty body on two fetch attempts this run (2026-10-11), so the repository README stands in as the primary source.

Strengths
#

  • The review gate is structural: an assignee cannot accept its own submission, and rejection feedback routes back to the exact agent session under a server-owned reservation that prevents duplicate sends.
  • Provenance is structural too: claims record the verified Realmroot binding’s runtime and session, humans observe the bound session without gaining runtime control, and public board views share progress without exposing the workspace.
  • Dependency-aware planning with cycle and cross-tenant rejection, on compare-and-swap writes rather than last-write-wins.
  • The product practices what this section preaches: behavior is specified in spec/*.feature files with traceability markers carried into tests.

Cautions
#

  • It is not self-contained: identity, runtime, and execution come from Realmroot and Enbor, so adopting the board means adopting an ecosystem, a deeper coupling than any other member of this category carries.
  • FSL-1.1 is source-available, not open source, and each release stays proprietary to competing use for two years before its Apache-2.0 conversion.
  • Four months without a tagged release at 488 stars is a thin public scorecard for a system of record; the commit stream is the only liveness evidence.
  • The name collides with an older VS Code Agent Kanban whose launch thread outranks it in every search, so discoverability is a standing tax.

Pricing
#

Free to deploy: the Worker runs on your own Cloudflare account (D1 plus configured secrets) under FSL-1.1 terms. No paid tier or hosted plan is documented.

Compared to
#

  • It’s a Plan: the closest comparable, a team board where agents hold slots; It’s a Plan is AGPL-3.0 and self-hosted on your own Postgres, Agent Kanban is FSL-licensed on Cloudflare’s D1 inside the Realmroot ecosystem.
  • beads: the repo-local counterpoint; beads needs no identity provider, while Agent Kanban’s claims are only as strong as the Realmroot bindings behind them.
  • Kata: the local-ledger counterpoint, machine-local and MIT versus cloud-native and ecosystem-bound.

Bottom line
#

Recommended for teams already inside the Realmroot ecosystem, or any team that wants review enforcement stronger than prompt conventions and accepts Cloudflare plus two external services as the price. Not for repo-local workflows, air-gapped environments, or anyone who needs an open-source license today. My disagreeable claim: making the assignee unable to close its own task is the most copy-worthy rule in this category, and it is telling that the one product enforcing it is the least standalone.

Changes
#

  • 2026-10-11 - Created after the topic search surfaced the enforced-review design, with the release gap and the name collision recorded.
  • 2026-10-11 - Added the star history chart.

See also
#

References
#