dsh-algovault

Mount the AlgoVault MCP server in DeepSeek Harness with one command.

This bundle ships a preconfigured @deepseek-ai/dsh-mcp-client row pointed at https://api.algovault.com/mcp. Your agent gets composite BUY / SELL / HOLD trade calls, market regime, cross-venue funding arbitrage and the live track record, as native tools.

Built by AlgoVault Labs — algovault.com

Install

dsh plugin --profile <name> add github:AlgoVaultLabs/dsh-algovault

Then restart that profile. Bundle membership is read at start, not hot-reloaded.

From the dsh.pub registry, the pinned form is:

npx dshpub add AlgoVaultLabs/dsh-algovault --ref <commit>

There is no build step and no key to configure. pnpm must be on PATH; dsh plugin forwards to it.

Tools

Every tool arrives namespaced as mcp__algovault__<tool>.

The bundle also ships a skill at skills/algovault-verdicts/SKILL.md that teaches the model which tool answers which question. Copy it into ~/.dsh/skills/ to load it.

Tiers

The free tier is anonymous. Install, restart, call — no key, no signup.

Paid tiers raise the quota and unlock the full funding-arb result set. Add the header in your profile's own cordis.patch.yml, not here, so an update to this bundle never overwrites your key:

- id: mcp-algovault
  name: '@deepseek-ai/dsh-mcp-client'
  config:
    serverName: algovault
    transport: streamable-http
    url: https://api.algovault.com/mcp?src=dsh-bundle
    headers:
      Authorization: !!js `Bearer ${process.env.ALGOVAULT_API_KEY}`

Set ALGOVAULT_API_KEY to your key, which looks like av_live_.... A patch replaces the whole config, so restate every field above, not only headers.

Current quotas and tiers: api.algovault.com/signup.

Model Experience

The model sees the tools above under the mcp__algovault__ prefix. Each returns a structured verdict rather than raw indicator values, so the model reads a decision and its confidence instead of assembling one.

A verdict of HOLD is a real answer, not a failure. The model should report it and stop, rather than retrying with different parameters until a directional call appears. Confidence and market regime belong in the reply beside every verdict; a BUY in a VOLATILE regime is a weaker claim than a BUY in a trending one.

AlgoVault supplies the thesis. It places no orders and holds no funds. The model should never present a verdict as an instruction to execute.

Win rates and coverage figures change. The model should quote them from a get_track_record response, never from memory.

Known Limitations

DeepSeek Harness is a developer preview and its own README warns of compatibility-breaking changes. Every published version is a release candidate. This bundle is deliberately thin for that reason: one client row, no wrappers around harness internals. Verified against @deepseek-ai/dsh@0.1.1-rc.2 and @deepseek-ai/dsh-mcp-client@0.1.1-rc.2 on 2026-08-30.

The bundled skill is not auto-discovered. The harness scans project, custom and user skill roots, and a bundle's own directory is none of those, so the copy step above is required.

MCP resources and prompts are not bridged by the harness. Tools only.

The endpoint is a hosted HTTP service. If it is unreachable at startup the harness still boots and logs an error, and the AlgoVault tools are absent for that session.

Links

• Integration guide
• Track record
• Issues

License

MIT. See LICENSE.