dsh-deepbrain

Bring DeepBrain (深脑) into DeepSeek Harness.

DeepBrain does not store documents — it stores judgments: conclusions distilled from recordings of meetings, interviews and lectures, each with an evidence chain back to what was actually said, and corroborated across separate occasions. With this plugin your DSH agent can read those judgments, quote verbatim-verified lines under someone's name, look up a person's historical positions, and borrow DeepBrain's analysis methods for material of its own.

Requires a DeepBrain account and API key. DeepBrain: https://shennao.zaowuyun.com The product UI and returned content are in Chinese.

Install

npx -p @deepseek-ai/dsh dsh plugin --profile web add github:qiuyiwu1989-star/dsh-deepbrain

Build output ships in the repository, so no dependency build scripts need to be allowed. Restart the profile afterwards.

Then set your key (generate one under Settings → API keys; it looks like lj_live_…):

export DEEPBRAIN_API_KEY=lj_live_xxx

Prefer the environment variable over putting it in config: profile config is printed verbatim by dsh --dump-config, so a key written there ends up in logs and screenshots.

Related MCP server: DeepAnalyze MCP Server

Six tools

The server currently exposes 27 MCP tools. This plugin registers 6 — a client's job is not to mirror the server, it is to present the right shape. Hand a model 27 tool names and its default strategy becomes "use whichever returns the most", which means going straight for raw transcripts and blowing up its own context.

Leaving tool empty on deepbrain_call lists every tool the current key is allowed to call.

Configuration

- insert:
    - id: deepbrain
      name: 'dsh-deepbrain'
      config:
        endpoint: 'https://shennao.zaowuyun.com/api/mcp'   # self-hosted? change this
        apiKey: ''                                          # empty → read DEEPBRAIN_API_KEY
        timeoutMs: 120000                                   # multi-step retrieval can take a while
        passthrough: true                                   # false = don't register deepbrain_call

Scopes

DeepBrain API keys are scoped, and the server trims the callable tool set accordingly:

• brain.ask — judgments, quotes, people, method library. Most usage needs only this

• transcript.read / transcript.write — fetch raw transcripts, feed new ones in

• brain.propose — write judgments back (they land in an inbox and require human review)

When a scope is missing the tool tells the model plainly that retrying will not help and a different key is needed, rather than reporting a generic failure.

What it does not do

• No caching. Every call is live. Judgments get superseded and expire; a cache would serve stale conclusions as current ones.

• No writes unless you explicitly call a write tool through deepbrain_call, which requires brain.propose.

• No web UI. Host-only plugin: it contributes tools, registers no slots, ships no client bundle.

Known boundary

Verified end-to-end in a real profile on 2026-09-07: installed from GitHub, loaded, called production with a real key, got real content back; a scope-limited key correctly returns "use a different key" rather than a generic failure. The note below only applies to a bare profile without auto-install-peers.

@deepseek-ai/dsh-tools is a real dependency here (installed with the plugin), but it declares @deepseek-ai/cordis as its own peer. In a clean profile without auto-install-peers, cordis will not be installed — this is upstream of this plugin and affects any third-party plugin that uses dsh-tools.

Normally the harness supplies cordis (it is a dependency of dsh itself). If you hit Cannot find package '@deepseek-ai/cordis', run this in the profile:

pnpm config set auto-install-peers true

This plugin deliberately does not depend on its own copy of cordis — cordis is the DI container, and duplicating it splits Context/Service identity, which is far worse than a missing dependency.

Development

See CONTRIBUTING.md.

MIT