Products

MCP

The full Hosho review inline in Cursor, Claude Code or any MCP client.

What it is

One MCP server that adds the Hosho review tools to Cursor, Claude Code or any MCP client. Ask your agent "Hosho review this prompt" and the full review - scores, line-anchored findings, rewritten sections - arrives inline in the IDE.

Tool	What it does
`hosho_improve_prompt`	Reviews a system or agent prompt and returns macro-factor scores (Scope, Structure, Guidance, Coherence) broken into sub-factors, specific findings tied to lines, an overall verdict, and rewritten sections for weak areas.
`hosho_compare_prompts`	Compares two versions of a prompt and classifies each change as an improvement or a regression, with severity and suggested fixes for the regressions.
`hosho_check_skill_alignment`	Audits a set of documents - a main prompt plus the skills and sub-prompts it invokes - and flags duplication, contradictions and dangling references between them.

With co-presence installed, the same server also carries the live-session tools (join, chat, highlight, propose) - see that page.

Install - just you

Add this to your MCP client config - ~/.cursor/mcp.json, or VS Code's MCP settings. The pfr_... value is your API key - mint one at /setup/mcp.

{
  "mcpServers": {
    "hosho-prompt-review": {
      "url": "https://hosho-mcp-server.onrender.com/mcp",
      "headers": {
        "Authorization": "Bearer pfr_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Already installed the co-presence plugin? The MCP came with it - nothing to do here, and doing both registers every tool twice under two names.

Install - your whole team

Commit this as .mcp.json at your repo root - everyone who clones the repo gets the tools. The key lives in each person's shell as HOSHO_API_KEY, not in the file.

{
  "mcpServers": {
    "hosho-prompt-review": {
      "type": "http",
      "url": "https://hosho-mcp-server.onrender.com/mcp",
      "timeout": 900000,
      "headers": {
        "Authorization": "Bearer ${HOSHO_API_KEY}"
      }
    }
  }
}

The timeout is not optional: a full Hosho review routinely runs 5-10 minutes, while most MCP clients default their per-request timer to 60 seconds. 900000 ms keeps the client listening for the whole review.

What a review returns

The overall score out of 10 and a verdict.
The four macro factors with their sub-factors - each a percentage of checks passing, so you can see where the prompt is weak at a glance.
Findings tied to specific lines, each with the problematic text and a suggested fix.
A link to the full report as a shareable page.

Before reviewing, the agent asks two questions: which model family the prompt targets (Claude, GPT, Gemini, DeepSeek, Qwen, Kimi or GLM) and whether that model is a standard or a reasoning model. The answers tune the provider-fit checks - a reasoning model is judged against different conventions than a standard one. In CI the same choice is pinned deterministically with a models.md file - see the Git Bot page.

Reviewing a component in its runtime context

A skill or sub-prompt rarely runs alone - it runs inside a larger assembled prompt. Hosho can review one file as it actually runs: hand it the parent prompt, the sibling skills and the shared references, and the findings stay scoped to the file under review while judging it in context - so a skill is not marked "missing a persona" when the parent supplies one.

Re-reviews remember your decisions

Hosho will not argue with a decision you already made.

On a re-review, findings you accepted or dismissed are not re-recommended, and a fix you implemented by hand - without formally accepting the card - is detected as resolved. This is what stops the circular loop where fixing one finding resurfaces as a new one on the next pass.