dsh-mcp-bridge

Hands a DeepSeek Harness deployment's capabilities to any MCP client — memory, the durable task ledger, project-agent dispatch and the wider MCP estate — as declared tools instead of shell commands a model has to be told about and remember.

Why

claude -p runs its own agent loop, so it never receives the harness's tool schemas: a Claude-backed session has no memory, no ledger, no dispatch. It does accept MCP servers. This is that server.

Everything it exposes already has a command-line entry point, so this adds a surface, not authority — it can do exactly what the caller could already do in a shell. What it cannot bridge is anything that only exists inside a live harness turn (goals, AgentTeams); those are session-scoped by design and have no external surface.

Tools

tool	does
dsh_memory_search / dsh_memory_write	the durable memory store
dsh_tasks_list / _add / _done / _block	the task ledger
dsh_dispatch / dsh_job_status	hand work to a project agent, then poll it
dsh_mcp_call / dsh_mcp_schema	any server in the MCP estate, uncomposed
dsh_projects	the repository ownership registry
dsh_notify	reach the user

Use

// mcp-bridge.json
{ "mcpServers": { "dsh": { "command": "node", "args": ["/path/to/src/server.js"] } } }

claude -p --strict-mcp-config --mcp-config ./mcp-bridge.json

Point DSH_JARVIS_BIN at the directory holding the CLIs if it is not ~/.dsh-jarvis/bin.

Three things learned the hard way

• --mcp-config is variadic. Inline JSON followed by a positional prompt makes the client read the prompt as a second config path (MCP config file not found: /tmp/List the tool…). Pass a file.
• Don't exit the instant stdin reports EOF. The spec says client-closes-stdin is shutdown, but some clients hand the server a stdin already at EOF; the server exits before the first request, the client respawns it, and it loops. Observed: 32 spawns in three minutes, no call completed. Here EOF only ends the process once something has been served, with a bounded idle timeout so an abandoned server still goes away.
• Verify a bridge with a write, never a read. A model asked to "call the search tool" can reproduce plausible output from its own context when the call actually failed — one did exactly that here and later admitted it. A write with a random token either lands in the store or does not.

Tests

npm test

20 tests, no dependencies, no client required: the tool catalogue and the JSON-RPC layer are both pure.

License

MIT