@agentproto/adapter-mastra-agent

The first-party agentproto agent. Every other adapter (codex, hermes, claude-code, opencode) wraps an external agent CLI. This one is ours end to end: an AIP-42 AGENT.md is run as a live Mastra agent behind an AIP-44 ACP server — our loop, our models, no third-party CLI.

# Standalone — drive it from any ACP-speaking host over stdio:
agentproto-mastra acp --model anthropic/claude-opus-4-8

# Or let the agentproto daemon spawn it like any other arm:
#   agent_start({ adapter: "mastra-agent", model: "openrouter/z-ai/glm-5.2" })

How it works

AGENT.md ──parseAgentManifest──▶ AgentHandle
                                     │ buildMastraAgent (@agentproto/mastra)
                                     ▼
                              Mastra Agent.stream()
                                     │ text deltas
                                     ▼
          MastraAcpAgent (src/acp-host.ts)  ──agent_message_chunk──▶ ACP client
                                     ▲
                       @agentclientprotocol/sdk AgentSideConnection (stdio)

• Model — any provider/model string Mastra's gateway can route (anthropic/…, openrouter/…, openai/…). The provider key is read from the spawn environment. Default: openrouter/z-ai/glm-5.2.
• Agent — pass --agent ./path/AGENT.md (or AGENTPROTO_MASTRA_AGENT_FILE) to run a custom agent; omit for a built-in coding default.

Workspace tools

The default agent is granted a workspace toolset, all confined to the session cwd (path-traversal guarded):

| Tool | Does | | --- | --- | | list_dir | List a directory. | | read_file | Read a UTF-8 file. | | write_file | Create/overwrite a file (mkdir -p). | | edit_file | Replace a unique substring. | | run_command | Run a shell command (cwd-scoped, timeout). |

Note: read_file and write_file here are Mastra-native workspace tools scoped to this adapter's session cwd. They are distinct from the runtime MCP filesystem tools (file_read, file_write) registered on the daemon's /mcp endpoint.

run_command is on by default; set AGENTPROTO_MASTRA_NO_EXEC=1 to withhold it.

Memory

Per-conversation memory via Mastra's LibSQL (SQLite) store. Each ACP session is a memory thread, so the agent recalls earlier turns within a session. The db is a single SQLite file under ~/.agentproto/mastra-agent/memory.db (persistent across spawns), overridable with AGENTPROTO_MASTRA_MEMORY_DB. A custom AGENT.md tunes it via the memory: block (scope, retention_turns).

Status

Streaming conversation + workspace tools (edit/run) + SQLite memory, with tool activity surfaced live: each tool the agent runs is relayed to the host as an ACP tool_call (status in_progress) and then a tool_call_update (status completed/failed with the raw output) — read off Mastra's fullStream and mapped in src/tool-call-map.ts. So a host (codex/claude-code/an IDE) shows the "🔧 run_command: …" feed, not just the final prose.