Excalibur

Agents that only write code leave you managing the bookends yourself — scoping what to build, verifying it works, documenting it honestly, and auditing what shipped. Excalibur covers that full cycle: Discovery analyzes scope, evidence, and risk before producing a scoped work item (with a recommendation of build, validate, or don’t-build), then Plan → Build → Test → Docs → Review → Ship → Audit runs as a structured pipeline. Large tasks auto-size into a swarm of agents running in isolated worktrees, coordinated through a verification mesh that gates completion on typed claims like tests_passed, type_safe, and no_secrets.

The defining technical feature is the immutable, append-only event log. The vendor describes this as a single stream that serves simultaneously as the live terminal view, replay scrubber, browser dashboard, and audit trail — byte-identical across all four surfaces. You can rewind mid-session with Esc-Esc, fork a new run from any cached step (replaying the good prefix for free), or mint a read-only share link from the local dashboard. The docs describe this as architecturally distinct from other agents, which reconstruct state rather than preserving it.

Excalibur is self-hosted, Apache-2.0 licensed, and installs via a single npm command. It is provider-agnostic — the docs describe support for OpenAI-compatible endpoints (including Azure), Anthropic, Ollama, and others, with model switching mid-session via /models. Keys live in an env file and never touch the repo. The local board (excalibur serve) runs as one process on localhost, giving teams Kanban work items, live cost tracking, and swarm chronograms without creating an account. The tool is positioned for product engineers and teams that need end-to-end traceability on a local repository — not for teams that need a hosted API surface or plug-and-play CI integration.