MCP Lens v0.1.0-rc.2: 1,000 MCP tools, two model-facing schemas

MCP Lens is a progressive-disclosure MCP gateway for DeepSeek Harness.

• mcp_search finds capabilities and reveals exact selected schemas.
• mcp_call invokes one exact server/tool.
• Connections are lazy; catalogs are last-good and server failures stay isolated.
• Search and direct calls share a default-deny allow/deny policy.
• Discovery, pagination, tools, catalogs, cursors and HTTP responses have explicit bounds.

Remote tools	Official client schema JSON	MCP Lens	Reduction
12	4,862 B	1,114 B	77.088%
100	62,062 B	1,114 B	98.205%
1,000	647,962 B	1,114 B	99.828%

Metric: UTF-8 bytes of Harness-visible serialized tool schemas. This is not tokenizer-token, billing-cost or LLM task-quality evidence.

Verified for this release: Node 22.19+ / 24+, 58 tests, real stdio and Streamable HTTP fixtures, production dependency audit, deterministic tarball rebuild, and secret/forbidden-file scans.

Install the attached prebuilt tarball or use the pinned Git source command from the README.

MCP Lens is an independent community plugin, not a sandbox and not an official DeepSeek component.

dsh-mcp-lens 0.1.0-rc.1 public refresh

dsh-mcp-lens is a progressive-disclosure MCP gateway for DeepSeek Harness.

This public refresh keeps the model-facing MCP surface fixed at two tools:

• mcp_search
• mcp_call

Measured on the same Harness serialization path:

• 12 tools: 4,862 -> 1,114 bytes (-77.088%)
• 100 tools: 62,062 -> 1,114 bytes (-98.205%)
• 1,000 tools: 647,962 -> 1,114 bytes (-99.828%)

Included in this release:

• release tarball
• reproducible benchmark artifact
• hardened .gitignore, SECURITY.md, GitHub Actions verify workflow, Dependabot
• public hero asset and launch copy

Boundaries:

• schema bytes are not tokenizer billing
• retrieval fixture is not a causal user-quality benchmark
• this plugin does not add OAuth, resources, prompts, or MCP task execution

MCP Lens v0.1.0-rc.1: 1,000 MCP tools, two model-facing schemas

What shipped

dsh-mcp-lens adds a progressive-disclosure MCP layer to DeepSeek Harness.

• mcp_search: search remote MCP capabilities and reveal only the selected exact schemas
• mcp_call: invoke one exact server/tool
• lazy connections, TTL-bound last-good catalog, and failure isolation
• consistent allowTools / denyTools policy on both search and direct call
• bounded discovery, bounded HTTP response bytes, and credential-aware cache persistence

Measured result

Exact visible tool-schema surface:

• 12 remote tools: 4,862 -> 1,114 bytes (-77.088%)
• 100 remote tools: 62,062 -> 1,114 bytes (-98.205%)
• 1,000 remote tools: 647,962 -> 1,114 bytes (-99.828%)

The metric is exact UTF-8 JSON bytes of ctx.tools.schemas(). It is not tokenizer tokens, provider billing, or task-quality evidence.

Verified on

• DSH 0.1.0-rc.6
• Node 22.19+ and 24+
• local verify: 58/58 tests passed
• npm audit --omit=dev: 0 vulnerabilities
• secret scan: detect-secrets found no results

Install

dsh plugin --profile web add github:labmimors/dsh-mcp-lens#v0.1.0-rc.1

Or install the attached reviewed release tarball.

Boundary

This plugin is not a sandbox. It reduces standing schema exposure; it does not make an untrusted MCP server safe.