Listspace

Listspace for coding agents

The pieces that let an agent use a Listspace board as its project memory. This folder is public as github.com/listspace/agent, and the CLI is on npm as listspace. Install and sign in with:

npx listspace login     # paste a personal token from Listspace: Settings > Your AI > API tokens

npx listspace boards

The MCP server is also listed in the MCP Registry as app.listspace/listspace and on Smithery.

Editing the skill

Edit skills/listspace/ only, then run:

node agent/scripts/sync-skill.mjs

node agent/scripts/check-agent.mjs

The plugin gets a real copy, not a symlink, so the plugin folder still works when it is copied or downloaded on its own. CI fails when the two differ. Keep SKILL.md short: steps and rules there, detail in reference/.

One version runs through cli/package.json, cli/src/version.ts, plugin.json and the skill's metadata.version; the check fails when they differ.

Try it locally

The CLI

cd agent/cli

npm ci npm run build node dist/listspace.mjs login # paste a token from Listspace: Settings > Your AI > API tokens node dist/listspace.mjs whoami node dist/listspace.mjs session " " node dist/listspace.mjs add " " ideas "Try dark mode" --description "Seen in settings"

npm link in agent/cli puts listspace on the PATH. LISTSPACE_TOKEN overrides the saved token for one run, and LISTSPACE_API_URL and LISTSPACE_MCP_URL point the CLI at another server. Every command takes --json. The token is saved in the OS config directory (%APPDATA%\listspace on Windows, ~/Library/Application Support/listspace on macOS, ~/.config/listspace elsewhere) as credentials.json, mode 600.

Tests: npm test (node:test with a fake fetch, no network). npm run typecheck runs the strict type check.

The skill on its own

Any agent that reads Agent Skills:

npx skills add ./agent -s listspace     # the skills CLI reads skills/ in a local folder

or copy agent/skills/listspace to ~/.claude/skills/listspace for Claude Code. Connect the MCP server separately (claude mcp add --transport http listspace https://mcp.listspace.app), or use the CLI above.

The Claude Code plugin

For one session:

claude --plugin-dir ./agent/claude-code-plugin

Or as a local marketplace, which keeps it installed:

claude plugin marketplace add ./agent

claude plugin install listspace@listspace

Then run /mcp, pick the listspace server and sign in (OAuth through Listspace). The skill shows up as /listspace:listspace and Claude loads it on its own when the project has a board. claude plugin validate ./agent/claude-code-plugin checks the manifest.

If you already added the Listspace MCP server yourself, Claude Code may show both; remove one.

The decision check

The plugin also has a small check that runs at the end of each of Claude's turns (hooks/hooks.json, hooks/decision-check.mjs). When the user picks an option, says yes to a proposal or corrects a fact, and Claude records no decision in that turn, Claude is sent back once with one line: if it was a decision that affects future work, record it with record_decision as Proposed; otherwise end the turn without comment.

The rules lean to silence. A yes to an offer to act ("Want me to commit?" "yes"), a maybe, and a bug report or a place in the code ("no, it's still failing", "the bug is in the parser") are not decisions. A yes counts only when Claude's last message offered a choice or made a proposal (recommend, suggest, propose, options, which one, a lettered list).

• It does nothing in a session that has not used a Listspace tool, and never blocks twice in a turn (stop_hook_active), so it cannot loop.
• It is plain rules on the user's words, with no model call and no network, so it costs nothing and answers in milliseconds (timeout 5 seconds).
• It needs Node.js 18 or later on the computer (Windows, macOS or Linux). Each command ends in ; exit 0, so without Node.js the check does nothing and Claude Code shows no hook error (sh, Git Bash and PowerShell).
• It keeps a few small files per session in the plugin's data folder (CLAUDE_PLUGIN_DATA, readable only by the user) and removes them after a week. An ordinary prompt writes nothing. A prompt that reads like a decision leaves its first line, cut at 80 characters, for that turn only. Of Claude's last message it keeps only whether it offered a choice. Without the data folder the check does nothing.
• Tests: node --test agent/scripts/decision-check.test.mjs. check-agent.mjs checks hooks.json.

The progress check

A second check at the end of each turn (hooks/progress-check.mjs) keeps the board in step with the work. When a turn changed files (Edit, Write, MultiEdit, NotebookEdit), committed, pushed or merged with git, opened or merged a pull request with gh, or deployed, and Claude made no Listspace write after that work, Claude is sent back once with one line: "Update the Listspace board: what did you start or finish this turn? Move the item, add a short comment."

• It does nothing in a session that has not used a Listspace tool, and never blocks twice in a row (stop_hook_active), so it cannot loop.
• After a nudge it waits at least 3 turns before the next one, and only nudges again for new work, so it does not nag.
• Any Listspace write counts as the update (move_item, add_comment, complete_item, record_decision, capture and the rest); reads do not.
• It keeps only the facts (work done, board updated, the turn of the last nudge) in the plugin's data folder, never a command or a file name, and removes them after a week. Same Node.js and ; exit 0 rules as the decision check.
• Tests: node --test agent/scripts/progress-check.test.mjs.

Board context

The board rules arrive with the Listspace server's instructions at the start of a session, and a long session, most of all one whose conversation was compacted, can lose them. hooks/board-hooks.mjs adds them back to Claude's context at the moments that matter. It only adds context: it never blocks anything and never changes a command or its output.

• Session start (SessionStart: startup, resume, clear): one line that names the project's board and says to start with get_account and start_session, update the items in unfinished_from_last_sessions first, and keep the board current.
• After a compaction (SessionStart with source compact): the board rule, and to call get_account (or start_session) before going on, since claimed_by_you and started_by_you list the items Claude was working on.
• After a pull request or a deploy (PostToolUse on Bash and PowerShell): after gh pr create, gh pr merge or a deploy command that worked, one line says to move the matching item (to Doing for a new pull request, to Done after a merge or a deploy) and add a short comment.
• It does this only in a project whose CLAUDE.md or AGENTS.md has a Listspace board: line (in the folder or one above it, up to the repository root), or in a session that used a Listspace tool (the progress check's note). From the file it takes only a link to a Listspace board, never other text.
• It writes nothing and makes no network call. Same Node.js and ; exit 0 rules as the decision check.
• Tests: node --test agent/scripts/board-hooks.test.mjs.

What is left out, and why

• Device pairing for `login`: the API has it (POST /auth/device), but the CLI does not use it yet. login already separates "how the token is obtained" from saving it; a TODO in cli/src/cli.ts marks where pairing goes. --token and a piped token stay for scripts.
• A token in the plugin: .mcp.json has no headers on purpose. A token would ship to every user; sign-in through /mcp is per user. The check script refuses headers there.
• The CLI inside the plugin (bin/): left out, because claude.ai and Cowork do not install plugins with a top-level bin/ and the MCP server covers Claude Code.

Publishing

1. Skill via `npx skills add`: published. The public repository listspace/agent holds this folder at its root, so npx skills add listspace/agent works.
2. CLI on npm: published as the unscoped listspace (0.6.0). To release a new version: bump the one version (see "Editing the skill"), check the package contents with npm pack --dry-run, then npm publish from agent/cli with an npm account with 2FA. prepublishOnly runs the type check, tests and build.
3. Claude Code plugin: the public repo doubles as a marketplace (.claude-plugin/marketplace.json), so users run /plugin marketplace add listspace/agent and /plugin install listspace@listspace. Listing in Anthropic's plugin directory is a separate submission; the owner decides on it, with privacy policy and terms URLs for that listing.

Sources: Claude Code plugin manifest reference (https://code.claude.com/docs/en/plugins-reference), plugin loading and --plugin-dir (https://code.claude.com/docs/en/plugins/loading), marketplace reference (https://code.claude.com/docs/en/plugins/marketplace-reference), skills (https://code.claude.com/docs/en/skills), Agent Skills specification (https://agentskills.io/specification), the skills CLI (https://github.com/vercel-labs/skills), and Postiz's agent repo as the model (https://github.com/gitroomhq/postiz-agent).