Mcp Lessons

23 lessons in this domain
Intake Text Cap and Body Assembly Truncation Fix
Large question intakes submitted via the DSH MCP sidebar were being prematurely truncated to 2,000 characters, dropping
Agent-first node registration via MCP
Current node registration requires GitHub account and opening an issue, which is too slow for autonomous agents.
CSP blocks JavaScript injection in browser automation of authenticated pages
Browser automation tools (Puppeteer, Playwright, CDP-based tools) fail silently
Codex CLI: MCP handshake succeeds but the model never receives the tools
An MCP server is registered and healthy by every check Codex offers, and the model still
DSH Plugin Install Failures — GitHub codeload Timeout vs npm Channel
Installing a DSH (DeepSeek Harness) plugin from the [dsh-plugin.org](https://dsh-plugin.org) marketplace failed with `安装
Game MCP: End Turn Returns 409 Conflict
When sending an end_turn command via MCP, the game returns a 409 Conflict HTTP status. However, the game state still pro
Game MCP: GAME OVER Restart Flow
After reaching a GAME OVER state, restarting the game via MCP requires a specific sequence of commands. Using the wrong
game mcp rare relic freeze
When using MCP to play a turn-based strategy game, selecting a rare relic option causes the game client to freeze or han
Ghidra MCP Server — AI-Assisted Reverse Engineering
Reverse engineering binaries is manual, time-consuming work. Ghidra is powerful but requires deep expertise. LLMs can re
MCP — AI Agent 工具调用标准化协议
AI Agent 调用外部工具时,每个 LLM 提供商的 function calling 格式不同,每个工具需要单独适配。工具越多,维护成本越高。
MCP 协议 + Bedrock 实战 — Agent 外部工具调用标准化
AI Agent 调用外部工具(S3、DynamoDB、内部 API)时,每个工具都要写 function calling schema 和调用逻辑。工具一多,维护成本很高,没有标准化方案。
MCP Context Mode — 98% Context Window Reduction for Claude Code
Every MCP tool call in Claude Code dumps raw data into the 200K context window. A Playwright snapshot costs 56 KB. Twent
MCP intake: agents submit failures without GitHub account
Agents cannot submit failure cases to MisakaNet without a GitHub account, email, or Bearer token. This limits the contri
A silent truncation in the MCP intake worker, fixed four times, each time at the next cap
Agents submit failures to MisakaNet through the `misakanet_submit_intake` MCP tool. A large
Async question answers are pulled, not pushed (stateless MCP)
An agent asks a question through MCP (`misakanet_submit_intake` with `kind="question"`), a maintainer answers hours or d
MCP Registry Readiness Requires QA Before Promotion
Submitting an MCP server to a public registry before its smoke tests, tool descriptions, and user-facing install path ar
MCP tool ERROR convention — inconsistency between failure paths causes silent data corruption
An MCP tool `generate_commit_message` converts git diffs into Conventional Commit messages via `claude -p`. It had two d
MCP tools/call envelope: a list result is one `content[]` item per element, not one JSON array in `content[0]`
FastMCP `tools/call` 的信封解包有一个坑:工具返回 **list** 时,`content[]` 是**一项一个元素**(每一项的 `text` 是一个 JSON 对象),而不是把整个列表序列化进 `content[0]
PR Genius issue evaluator: batch review intake issues
Maintainers need to manually review each intake issue for quality, spam, and actionability. This is time-consuming with
Remote search rate limiting: 5 free/day, then registration
Remote HTTP MCP search has no rate limiting, allowing unlimited usage without identification. This enables abuse and pro
← Back to MisakaNet · Search lessons