blackboard

A shared blackboard for multi-agent coordination. When a main agent spawns subagents, they start cold and can't talk to each other. This gives them one shared board — message channels plus shared key/value state — backed by a single SQLite file every agent process opens.

main:   post("work", "summarize the auth module", author="main")
sub:    read("work")                  → [{ content: "summarize the auth module" }]
sub:    post("results", "uses JWTs in src/auth/session.ts", author="sub")
main:   read("results", after_id)     → the answer, when it lands

Tools

• post(channel, content, author?) — append a message; returns its id.
• read(channel, after_id?, limit?) — messages newer than after_id (poll to wait for a reply).
• set_state(key, value) · get_state(key) — shared key/value (results, flags, counters).
• list_channels() — channels + message counts.

How it works

One SQLite file (default ~/.agent-blackboard/board.db, override with BLACKBOARD_DB) in WAL mode with a busy-timeout, so concurrent agent processes read and write the same board safely. node:sqlite — no native deps.

A typical pattern: main posts tasks to a channel, subagents read to pick them up and post results back; read(channel, after_id) polls for the reply.

Setup

npm install   # Node 22+ (built-in node:sqlite)
npm test

claude mcp add blackboard -e BLACKBOARD_DB=/tmp/run123.db -- node /abs/path/blackboard/dist/server.js
# CLI: agent-board post work "task" --author main | read work --after 0 | set k v | get k | channels

Honest limits

• No auth/ACLs — any process pointed at the file can read/write. Scope a run with its own BLACKBOARD_DB if you want isolation.
• Poll-based, not push — readers poll with after_id; there's no blocking wait.
• Messages accumulate until you rotate/delete the db (no auto-expiry).