# Distribution

How Bucker is listed where coding agents look. These are static artefacts; publishing any
of them is a human release step (nothing here publishes itself, and no scope joins any
ceiling: SG-059 is held).

| Artefact | Source | Published by |
|---|---|---|
| MCP Registry entry | `server.json` (name `io.bucker/bucker`, DNS-verified `io.bucker` namespace) | a person with `mcp-publisher`, after the `_mcp-registry` TXT record exists at Cloudflare |
| Server Card | `mcp-server/src/protocol/server-card.ts`, served at `https://mcp.bucker.io/.well-known/mcp/server-card` | bucker-mcp, on deploy |
| Claude Code plugin | `tooling/agent-plugins/claude-code/` | a person, through the Claude directory submission |
| Agent Plugins 1.0.0 (Codex, Cursor, Copilot, Kiro, VS Code), Cursor Marketplace, OpenAI directory | not built yet; same hook contract as the Claude Code plugin | a person |
| Client config | `bucker agent setup --client claude-code\|codex\|cursor\|vscode [--dry-run] [--skills]` | the developer, locally |

## Invariants

- The resource identifier is `https://mcp.bucker.io`: the origin, no path, no trailing slash. `server.json`'s remote URL and the card's `resource` use that spelling.
- The card lists `READ_SCOPES` only (the live ceiling); held scopes never appear. Its tool count comes from the registry, so every channel quotes the same number.
- Credentials are referenced by environment variable (`BUCKER_ACCESS_TOKEN`), never written into a file.
- The plugin hook answers only for the `bucker` server's tools, in both spellings Claude Code gives them: `mcp__bucker__*` (a project `.mcp.json`) and `mcp__plugin_bucker_bucker__*` (the plugin's own server). It asks for every Bucker tool that is not an exact name from the server's read table, and denies every Bucker call when Bucker cannot be reached. A crash or a hang of the script is exit status 2, Claude Code's blocking result, decided by a watchdog inside the manifest's timeout (fail closed). It never allows, and never answers for another tool; a wider hook is OD-DP-17's. The one case no script covers is a machine with no `node` to run it: the server's approvals engine is the control there, as for a client with no plugin.
- Claude Code loads a plugin from `.claude-plugin/plugin.json` at the plugin root; the person who packages `tooling/agent-plugins/claude-code/` for submission must put the manifest there (the repository keeps it beside its hooks, where the tests read it). Nothing packages it automatically yet.
- The `server.json` `version` tracks `SERVER_VERSION` in `mcp-server/src/server.ts`; bump them together. `mcp-server/test/server-card.test.ts` pins the version, the remote URL spelling and the card link.

## Directory listing text

Short: "Bucker: errors, traces and governed changes for coding agents. Reads are open to a signed-in agent; every write is a proposal a person approves in Bucker."
Long: the description in `server.json`, plus the documentation link `https://bucker.io/docs/mcp-setup`.
