@beeos-ai/bma-dsh-acp is the BMA build of the DeepSeek Harness (DSH) brain, served as an Agent Client Protocol (ACP) agent on stdin/stdout. The executor of your computer starts it and drives it: the brain thinks, and every tool it calls runs on your computer through the executor, so the hands are the same as for a platform brain. Sessions, events and approvals stay on BMA: you still send messages with /v1, and approvals reach every client of the session.

Install and run

  • Node.js 24 or later. Tested on Linux x64 and macOS arm64; Windows is untested.
  • BMA_DSH_ACP_HOME is required: the directory for the agent’s state (sessions are kept there for later loads). stdout carries ACP only; logs go to stderr.
  • DSH is not bundled. npm installs the official @deepseek-ai/* packages, whose versions @beeos-ai/bma-dsh-adapter pins exactly, including the transitive ranges DSH leaves open. One package version therefore always installs the same DSH (0.1.0 carries DSH 0.2.2-rc.1, reported by initialize as _meta.bma.dsh). Do not install with --omit=optional: the native file-lock addon is installed as an optional dependency for your platform.
  • Pin the version. npx @beeos-ai/bma-dsh-acp without a version takes the newest release; use @0.1.0 (the two packages @beeos-ai/bma-dsh-acp and @beeos-ai/bma-dsh-adapter always share one version). Upgrade by changing the pin.

Start it from the executor SDK

In the Go SDK (bma-executor-sdk 1.3.0 or later), ACPAgent starts the brain and serves BMA turns with it. The SDK never looks the command up on PATH, so Command[0] must be an absolute path:
npx and node are commands too: write their absolute paths (command -v npx), not bare names. The model endpoint comes from Upstream.

How it differs from the packaged forms

Sandboxes use the bundled form and need nothing from you. Use the npm form only on a computer you run an executor on.

Executor Protocol

Roles, connection and deployment shapes

Agent capability

Bring your own agent loop instead