MCP server de agent-board

agent-board

A self-hosted Kanban board for AI coding agents, with an MCP server built in.

English · Português

agent-board is an open-source task board that Claude Code, Codex and other AI agents read and write through the Model Context Protocol (MCP), while you use it in the browser like any Kanban. It gives every agent session the same task list and a history of who did what, and it runs on your own machine: one SQLite file, no cloud account.

When several AI sessions work on the same projects, each one starts from zero. None of them knows what the others did, what is in progress, or why a decision was made. This board is the shared state: you drag cards in the browser, and Claude Code, Codex and other agents read and write the same cards through MCP, a REST API or a small CLI. Every change is recorded with its author, so opening a card shows which session did what.

The interface is available in English and Portuguese (Settings → Customize → Language; it follows the browser language the first time). The API routes, the MCP tool names and the default column names are in Portuguese, because agents refer to them by name. The tool descriptions are self-explanatory to an AI agent; see CONTRIBUTING.md to add another language.

What it does

• Kanban with authorship. Columns, drag and drop, comments, and a per-card history that says who did each thing.
• For agents. 30+ MCP tools over HTTP, a REST API, and a dependency-free CLI. All three share one core, so they cannot drift apart.
• Finding things. Search across titles, descriptions, comments and tags (Ctrl+K), combinable filters kept in the URL, tags, and a project list with counts.
• Several boards, each with its own columns.
• Reminders and recurring tasks. Agents can ask for "reminders due today" when a session starts.
• Safe to operate. Archive instead of delete, undo (Ctrl+Z), and optimistic locking so two sessions do not overwrite each other.
• Three themes (dark, light, glass) and wallpapers stored on the server.
• English and Portuguese interface, switched in the settings.
• Small. SQLite in a single file, no external services, few dependencies.

Quick start

Requires Node.js 20 or newer. better-sqlite3 compiles native code on install; on Linux you may need build-essential and python3.

git clone https://github.com/Slayer1Dev/agent-board.git
cd agent-board

# Terminal 1: API and MCP on http://127.0.0.1:8078
cd server && npm install && npm run dev

# Terminal 2: interface on http://localhost:5174
cd web && npm install && npm run dev

Or with Docker (publishes both ports on localhost only):

docker compose up --build

Connecting an agent

MCP (Claude Code shown; any MCP client over HTTP works):

claude mcp add --scope user --transport http agent-board http://127.0.0.1:8078/mcp
# with a key:  --header "Authorization: Bearer YOUR_KEY"

Codex:

codex mcp add agent-board --url http://127.0.0.1:8078/mcp
# with a key:  --bearer-token-env-var BOARD_API_KEY

CLI, for agents without MCP or for your own terminal:

export AGENT_BOARD_URL=http://127.0.0.1:8078/api
export AGENT_BOARD_AUTHOR=claude

node cli/agent-board.mjs ver
node cli/agent-board.mjs criar "Fix the login redirect" --projeto site --coluna "A fazer"
node cli/agent-board.mjs mover 1a2b3c4d "Em andamento"
node cli/agent-board.mjs comentar 1a2b3c4d "Root cause was the cookie domain."
node cli/agent-board.mjs buscar "cookie"

A ready-made instruction file for Claude Code is in integrations/claude-code-skill. Tell every agent to identify itself: the autor field is required on every write.

The full list of routes and tools, with an example for each, is in docs/API.md.

Configuration

Variable	Default	Purpose
BOARD_DB	~/.agent-board/board.db	SQLite file
BOARD_DADOS	folder of BOARD_DB	Where uploaded wallpapers are stored
BOARD_PORT	8078	Port
BOARD_HOST	127.0.0.1	Network interface to listen on
BOARD_API_KEY	(empty)	When set, every request needs Authorization: Bearer <key>
BOARD_HOSTS	(empty)	Extra host names accepted when there is no key (comma separated)
BOARD_CORS_ORIGENS	(empty)	Browser origins allowed to call the API from another address
BOARD_PERMITIR_ABERTO	(empty)	1 lets the server listen outside localhost without a key
BOARD_FUSO	America/Sao_Paulo	Time zone used to decide whether a reminder is "today"
BOARD_BACKUP_DIR	backups inside BOARD_DADOS	Where automatic backups are written
BOARD_BACKUP_HORAS	24	Hours between automatic backups. 0 turns them off
BOARD_BACKUP_MANTER	14	How many backups to keep

Backups

The server backs itself up: no cron needed. Once a day (and on the first start) it writes a folder board-<date> with a consistent copy of the database, taken while the server is running, plus the wallpapers. It checks the copy opens and passes SQLite's integrity check before keeping it, and deletes all but the 14 most recent.

• Where: backups/ next to the database by default. Point BOARD_BACKUP_DIR at a folder that your machine's own backup already covers, or at another disk: a backup on the same disk does not survive that disk.
• See them: the settings menu shows the last one; node cli/agent-board.mjs backups lists them; agents have ver_backups.
• Make one now: "Back up now" in the settings menu, node cli/agent-board.mjs backup, or the fazer_backup tool. Worth doing before loading many cards at once.
• Restore: stop the server, copy board.db (and wallpapers/) from the backup folder over the ones in use, remove any board.db-wal and board.db-shm left beside the old file, start the server.

Security model

This is a personal tool, and its defaults assume a single machine.

• Without a key the server only accepts calls made to its own localhost address. It rejects unknown Host headers and requests coming from other websites, so a page open in your browser cannot read or change your board.
• To reach it from other machines, set BOARD_API_KEY. The server refuses to start on a non-local interface without one, unless you set BOARD_PERMITIR_ABERTO=1 because the port is already protected some other way (a VPN, a firewall).
• Do not put the key in the web build. Serve the interface through a proxy that adds the header: web/vite.proxy.example.mjs does exactly that.
• autor is declared, not verified. It is a history, not access control. Anyone with access to the API can write under any name.

See SECURITY.md to report a vulnerability.

How it is built

  web interface ──┐
  CLI ────────────┼──►  REST API ─┐
  Claude / Codex ─┴──►  MCP ──────┴──►  core (nucleo.ts)  ──►  SQLite
                                              │
                                    history with authorship

• One core. server/src/nucleo.ts holds all the rules; REST and MCP are thin layers over it.
• SQLite, not Postgres: one file, backup with cp, nothing to keep running.
• Stateless MCP. Each request creates its own transport; there is no session to expire.
• Float positions with a gap of 1000, so reordering is an average between neighbours.
• Native HTML5 drag and drop, no library.
• 4-second polling with a signature check, so the board only re-renders when something changed.

Known limitations

• Polling, not WebSocket: a change made by another session appears within 4 seconds.
• No user accounts. See the security model above.
• API routes, MCP tool names and server error messages are in Portuguese only. The interface translates the default board and column names for display, but agents still use the Portuguese names ("A fazer", "Revisão").
• Reminders are shown in the board and available to agents; nothing is pushed to phone or e-mail.

Tool names in English

The tool and route names are in Portuguese. An AI agent reads them without trouble; this table is for people.

Tool	What it does
contexto_sessao	Start a session here: what matters now, within a size budget
ver_quadro	View a board: columns and cards (compact by default)
ver_card	View one card: current state and last comments, or its paginated history
criar_card	Create a card
mover_card	Move a card to another column (how an agent reports progress)
atualizar_card	Edit title, description or project
comentar_card	Add a note to the card's history
arquivar_card / restaurar_card	Archive a card / bring it back
pesquisar_cards / filtrar_cards	Search / filter cards
atividade_recente	Recent activity: what changed and who changed it
lembretes_pendentes	Reminders that are overdue or due today
definir_lembrete / definir_repeticao	Set a reminder / make a task recurring
listar_quadros / criar_quadro	List boards / create a board
listar_tags / definir_tags	List tags / set a card's tags
ver_backups / fazer_backup	See backups / back up now

Default columns: A fazer (to do), Em andamento (in progress), Revisão (review), Concluído (done). The complete list is in docs/API.md.

FAQ

How do I make Claude Code and Codex share the same task list?

Run agent-board, add its MCP endpoint to both (see Connecting an agent), and tell each agent to read the board when a session starts and to move its card when it starts and finishes a task. Both now see the same cards, and each change is signed with the agent's name.

How is this different from Trello, Linear or GitHub Projects?

Those are built for people and reached by agents through an integration. agent-board is built the other way round: every operation an agent needs is an MCP tool, every write must say who made it, and the whole thing is one process and one SQLite file on your machine, with no account, no API quota and no data leaving your network. It has far fewer features than those products, on purpose.

How is this different from a TODO.md file in the repository?

A file works for one agent in one repository. agent-board covers several projects at once, keeps a history per card (who created, moved and commented), lets two sessions write at the same time without overwriting each other, and gives you a board to look at and drag.

Does it give AI agents memory between sessions?

It gives them shared, explicit memory of work: what is to do, what is in progress, what was decided and why, written in cards and comments that the next session reads. It does not store conversation transcripts and does no embedding or semantic search.

Which agents and clients work with it?

Any client that speaks MCP over HTTP (streamable HTTP). It is used daily with Claude Code and Codex. Agents without MCP can use the REST API or the CLI in cli/agent-board.mjs, which needs only Node.js.

Can I use it with a team or over the internet?

It is designed for one person and their agents on a private network. There are no user accounts: an API key protects the server, and the author of each change is declared, not verified. Read the security model before exposing it beyond localhost or a VPN.

What does it need to run?

Node.js 20 or newer, or Docker. No database server, no external service.

Contributing

Issues and pull requests are welcome. Read CONTRIBUTING.md first; if you work with an AI agent, point it to AGENTS.md.

License

MIT