ansur

Ansur — Claude Code plugin

Status: published. Distributed via the public marketplace repo ansur-ai/ansur-plugin — /plugin marketplace add ansur-ai/ansur-plugin then /plugin install ansur@ansur. This monorepo directory is the authoring source; the public repo is the published mirror (plugins/ansur/ there). Keep the two in sync when editing.

Turns Claude Code into the meta-harness — the build-time agent a customer pastes a prompt into, which drives the ansur CLI (@ansur-ai/cli) to build their AI employee on the Ansur platform.

The meta-harness (why this exists)

Two nested harnesses:

• Runtime harness — Ansur's own agent loop (TAOR) + pooled daemon + sandbox + guards. It runs the employee.
• Meta-harness — the build-time agent (Claude Code + this plugin). It authors the bundle the runtime harness executes. "Meta" = one level up: it doesn't do the employee's job, it produces the artifact (the bundle) that defines an agent that does.

Its only job is writing + iterating bundles — never platform code. A plugin (vs. MCP) ships behavior (a skill + commands) that shapes how the agent reasons, so the customer's agent behaves like Ansur's server-side meta-harness would — and the customer pays that inference. The CLI is the small action surface.

Background (monorepo paths — for plugin authors editing this source, not shipped to customers): docs/design-docs/platform-architecture.md (the meta-harness loop), docs/design-docs/platform-interface.md (plugin + CLI, why no web UI), docs/design-docs/core-beliefs.md (voice: real files/tools, context discipline, Oracle thinking).

Structure — one skill, a spine + on-demand depth by domain

The plugin ships a single skill (ansur) whose SKILL.md is a spine / lookup table of the flow. Depth lives in domain directories the spine points to and the agent reads on demand — not in separate registered skills.

plugins/claude-code/
  .claude-plugin/plugin.json   manifest
  commands/ansur-onboard.md    single slash entry point → loads the skill
  skills/ansur/
    SKILL.md                   the spine: mental model, build loop, hard rules (always loaded)
    setup/
      initial-setup.md         ONE-TIME: install + login + init + github connect
    bundle/                    the employee artifact
      bundle.md                the container: anatomy, manifest fields, create→push→reload lifecycle
      skills.md                on-demand playbooks the employee loads
      hooks.md                 bash gates on the employee's agent loop
      memory.md                what the employee accumulates across conversations
    guards/                    the safety boundary
      guards.md                connectors + wire guards: how the employee reaches external systems
      rules.md                 the rule LANGUAGE: expression grammar, when:/preflight:, authoring discipline
      sap.md                   the SAP connector recipe (two guard-systems, secrets, read role)
    operate/
      trace.md                 trace record shape + jq recipes

Encoded "don'ts"

• No ansur dispatch; no ansur bundle write/set-active (git is the versioning model).
• github is not a connector — ansur github connect is a control-plane credential.
• Never write guard code (the platform ships the guard); tools.yaml is vestigial.

Distribution

Published via a Claude Code plugin marketplace; the Track F3 install prompt installs this plugin + @ansur-ai/cli and runs ansur login. Codex CLI / Cursor variants reuse the same skill + CLI; only packaging differs.