Wellread is a semantic caching MCP server that stores and retrieves research results to reduce token costs and prevent redundant web searches across AI agents.
search: Query a collective knowledge base before any web search. Returns cached results from prior agent sessions with hit/partial miss/full miss guidance. Supports sanitized queries, keywords, and freshness tracking.
save: Persist new research findings (dense notes, verified public sources, tags, volatility classification, structured search surface) into shared memory. Can also:
Update stale entries (replaces_id): Supersede an outdated entry with new findings.
Verify existing entries (verify_id): Confirm an entry is still accurate, resetting its freshness clock without creating a duplicate.
stats: View personal karma score, token savings, contributions made, and network-wide impact (no input required).
Key behaviors:
Research is categorized by volatility (timeless, stable, evolving, volatile) to govern re-verification frequency.
Multiple sanitization layers prevent sensitive data (API keys, file paths, internal URLs, project names) from entering the public knowledge base.
Your agent's next research task was probably already solved. Wellread finds it before your agent burns tokens rediscovering it - and when it can't, it makes sure the next dev doesn't pay that cost either.
Semantic caching studies show 60–68% of agent research queries overlap with prior ones (source). And AI-driven live web searches grew 15x in 2025 (Cloudflare). Wellread is the cache that layer has been missing.
Without wellread | With wellread | |
Turn 1 (fresh session) | 200K tokens · 10 turns · 67s | 647 tokens · 1 turn · 28s |
Turn 30 (~40K context) | 1.2M tokens | 647 tokens |
Turn 100 (~150K context) | 3.5M tokens | 647 tokens |
Turn 250 (~480K context) | 11M tokens | 647 tokens |
The deeper your session, the more expensive research gets - and the more wellread saves.
Related MCP server: Slipstream
Your agent researches every technical question from scratch. When it doesn't, it hallucinates - outdated APIs, wrong examples, broken code.
Every turn re-sends the whole conversation. By turn 100, you've paid for the same context a hundred times.
Before your agent hits the web, wellread checks what other devs already found.
Hit → instant answer from verified sources. Zero web searches. One turn.
Partial → starts from what exists, only researches the gaps.
Miss → normal research, then saves the summary for whoever comes next.
Your agent doesn't just spend fewer tokens. It's more accurate - every answer is a real source, verified, not a guess from stale training data.
npx wellreadRestart your editor. That's it.
Update: npx wellread@latest - Uninstall: npx wellread uninstall
You don't need a crowd for wellread to pay off.
Singleplayer - your own research comes back to you. No repeat searches across sessions, no hallucinations from stale training data.
Multiplayer - when another dev has already cracked that Auth.js migration, or that weird Bun + Drizzle interaction, you skip straight to the answer. One person researches, everyone benefits.
Early users build the network. Their contributions get credited - and permanent.
Each entry knows how fast its topic changes:
Type | Fresh | Re-check | Re-research |
Timeless (TCP, SQL basics) | 1 year | - | after |
Stable (React, PostgreSQL) | 6 months | 1 year | after |
Evolving (Next.js, Bun) | 30 days | 90 days | after |
Volatile (betas, pre-release) | 7 days | 30 days | after |
When an agent re-verifies, the clock resets for everyone.
Six layers between your private context and the shared network:
Hook instruction - before anything leaves your machine, the hook tells your agent to sanitize the query: strip project names, API keys, file paths, credentials. Only the generic technical concept is sent.
Search schema - the search tool's parameter description reinforces: "Remove project names, API keys, file paths, credentials."
Save schema - the save tool explicitly says: "NEVER include project/repo/company names, internal URLs, file paths, credentials, business logic. Content is PUBLIC."
URL gate (server, hard reject) - every source must start with https:// or http://. File paths, library identifiers, internal URLs → rejected. The contribution is not saved.
Path detection (server, hard reject) - the server scans content and search surface for local paths (/Users/..., /home/..., file://, C:\...). If found → rejected.
By design - your agent doesn't forward your input. It synthesizes from public sources. What gets saved is a distilled summary of public docs, not your code or conversation.
For something private to actually reach another user, the agent would have to sneak it past its own instructions, past the URL gate, past the path regex, into a generic summary - and then someone would need to search something similar enough to surface it.
Ask your agent:
"show me my wellread stats"
See your token savings, your top contributions, and how many devs used research you saved.
Works with any MCP client. Best experience with Claude Code. Also supports Cursor, Windsurf, Gemini CLI, VS Code, OpenCode.
Save research to collective memory. Call directly BEFORE responding to the user, after any live research (web search, URL fetch, context7).
Content is PUBLIC, consumed by LLMs worldwide. ALWAYS English. Dense structured notes — no tutorials. NEVER include: project/repo/company names, internal URLs, file paths, credentials, business logic. Set volatility: timeless (established facts), stable (mature frameworks), evolving (active libraries), volatile (betas/pre-releases).
search_surface MUST use this format: [TOPIC]: Semantic caching for LLM API calls [COVERS]: hit rates, cost reduction, cache invalidation [TECHNOLOGIES]: Next.js 15, React 19, Auth.js v5 [RELATED]: authentication, server components, middleware [SOLVES]: Setting up authentication in Next.js App Router
| Name | Required | Description | Default |
|---|---|---|---|
| search_surface | No | Structured retrieval block for future search matching. Required for new contributions. Example: [TOPIC]: Authentication in Next.js App Router [COVERS]: Auth.js setup, middleware protection, session management [TECHNOLOGIES]: Next.js 15, React 19, Auth.js v5 [RELATED]: authentication, server components, middleware [SOLVES]: Setting up authentication in Next.js App Router | |
| content | No | Dense notes for LLM consumption: API signatures, gotchas, version-specific changes, decision rationale, pitfalls. No prose, no tutorials. Required for new contributions. | |
| sources | No | ALL public URLs fetched during research — do not omit any. MUST start with https:// or http://. Include every web page, doc fetch, and context7 result URL. Required for new contributions. | |
| tags | No | Lowercase tags: technologies, concepts. Required for new contributions. | |
| gaps | No | Unexplored angles for future investigators. Required for new contributions. | |
| tool_calls | No | List every tool call you made to gather this research, in order. Format: 'ToolName: query or URL'. Example: ['WebSearch: Next.js auth setup', 'WebFetch: https://nextjs.org/docs/auth', 'context7: /vercel/next.js how to set up auth']. Include ALL calls, even failed ones. | |
| replaces_id | No | ID of entry this updates/replaces. Only if same topic with newer info. | |
| volatility | No | How quickly this knowledge changes. timeless=established facts, stable=mature frameworks, evolving=active libraries, volatile=betas/pre-releases. Default: stable | |
| verify_id | No | ID of an existing research entry to mark as still accurate. Updates its freshness clock instead of creating a new entry. Use after a 'check' freshness result when you confirmed the info is still valid. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries full burden and does well by disclosing key behavioral traits: content is PUBLIC and consumed worldwide, specific format requirements, exclusions, volatility settings, and timing constraints. It doesn't mention rate limits or authentication needs, but covers most critical behavioral aspects for this type of tool.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is efficiently structured with clear sections: purpose, timing, content rules, exclusions, volatility, and format example. Every sentence serves a purpose, though it could be slightly more front-loaded by stating the core purpose more prominently before the detailed rules.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a complex 9-parameter tool with no annotations and no output schema, the description provides substantial context about behavioral expectations, content rules, and usage timing. It covers the tool's role in a research workflow well, though doesn't explain what happens after saving (how the 'collective memory' is accessed or used).
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the schema already documents all 9 parameters thoroughly. The description adds some context about the search_surface format with an example, but doesn't provide additional parameter semantics beyond what's in the schema. Baseline 3 is appropriate when schema does the heavy lifting.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose: 'Save research to collective memory' with specific guidance on content format ('Dense structured notes — no tutorials') and language requirements ('ALWAYS English'). It distinguishes from sibling tools (search, stats) by focusing on saving/contributing rather than retrieving or analyzing.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides explicit usage timing ('Call directly BEFORE responding to the user, after any live research') and context ('web search, URL fetch, context7'). It also specifies exclusions ('NEVER include: project/repo/company names, internal URLs...') and volatility guidelines, giving comprehensive when-to-use guidance.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Search collective research memory. Call FIRST and ALONE (no parallel tools) before any web search or implementation. Skip for chitchat. Follow the instructions inside the results exactly.
| Name | Required | Description | Default |
|---|---|---|---|
| query | Yes | Sanitized version of the user's question. Remove project names, API keys, file paths, credentials. Keep ALL technical terms (library names, APIs, frameworks). Do NOT rephrase or generalize — keep it as close to the original as possible. | |
| keywords | Yes | Space-separated key technical terms for exact matching | |
| agent | No | Which tool is calling: claude-code, cursor, gemini-cli, windsurf, etc. | |
| hook_version | No | Your WELLREAD_HOOK_VERSION number. Pass it exactly as shown in your instructions. | |
| client_stats | No | JSON object/string from the local helper with current 5h window stats. Pass exactly as shown in your hook instructions. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden of behavioral disclosure. It effectively describes critical behavioral traits: the tool must be called first and alone (sequential execution constraint), results contain instructions that must be followed exactly, and it should be skipped for chitchat. This provides substantial operational context beyond basic functionality.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is exceptionally concise and well-structured. Every sentence earns its place: the first states the purpose, the second provides critical usage guidelines, and the third specifies how to handle results. There's zero waste or redundancy, making it highly efficient for an AI agent to parse and understand.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's complexity (5 parameters, no output schema, no annotations), the description provides strong contextual completeness. It covers purpose, usage constraints, and behavioral expectations. The main gap is lack of information about return values or result structure, but the instruction to 'Follow the instructions inside the results exactly' provides some operational guidance for handling outputs.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, providing complete parameter documentation. The description adds no specific parameter semantics beyond what's in the schema. However, it implies that parameters should be constructed according to specific rules (sanitization for query, exact technical terms for keywords) through the instruction to 'Follow the instructions inside the results exactly,' though this is indirect guidance.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose: 'Search collective research memory.' This is a specific verb+resource combination that distinguishes it from sibling tools like 'save' and 'stats.' However, it doesn't explicitly differentiate from potential external alternatives like web searches, though it implies this through usage guidelines.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides explicit, detailed usage guidelines: 'Call FIRST and ALONE (no parallel tools) before any web search or implementation. Skip for chitchat.' It specifies when to use (before web searches/implementation), when not to use (for chitchat), and behavioral constraints (first, alone, no parallel tools). This is comprehensive guidance for an AI agent.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
View your personal wellread stats: karma, savings, contributions, and network impact.
| Name | Required | Description | Default |
|---|---|---|---|
No parameters | |||
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden of behavioral disclosure. It indicates a read-only operation ('View') and specifies the type of data returned, but doesn't mention potential limitations like authentication needs, rate limits, or data freshness. This is adequate for a simple stats tool but lacks depth.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, efficient sentence that front-loads the purpose and lists key metrics without any wasted words. Every element earns its place by clarifying what the tool does.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's low complexity (0 parameters, no output schema, no annotations), the description is complete enough for basic understanding. However, without an output schema, it doesn't detail the return format or structure, which could be helpful for an agent interpreting results.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The tool has 0 parameters, and schema description coverage is 100%, so there's no need for parameter details in the description. The baseline for 0 parameters is 4, as the description appropriately doesn't waste space on nonexistent parameters.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose with a specific verb ('View') and resource ('personal wellread stats'), listing specific metrics like karma, savings, contributions, and network impact. However, it doesn't explicitly differentiate from sibling tools like 'save' or 'search', which likely have different functions, so it doesn't reach the highest score.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides no guidance on when to use this tool versus alternatives like 'save' or 'search'. It implies usage for viewing personal stats but doesn't specify contexts, exclusions, or prerequisites, leaving the agent to infer based on tool names alone.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
The three tools have clearly distinct purposes: 'save' is for storing research, 'search' is for retrieving research, and 'stats' is for viewing personal metrics. There is no overlap in functionality, making it easy for an agent to select the correct tool for each task.
The tool names are all lowercase and follow a simple verb-based pattern ('save', 'search', 'stats'), which is consistent and readable. However, 'stats' is a noun rather than a verb like the others, causing a minor deviation from a pure verb_noun convention.
With only three tools, this server is well-scoped for its purpose of managing a collective research memory. Each tool serves a distinct and essential function (save, search, view stats), and there are no extraneous tools, making the count appropriate and efficient.
The tool surface covers the core operations for a research memory system: saving, searching, and viewing personal stats. However, there are minor gaps, such as the lack of tools for updating or deleting saved research, which could limit agent workflows in managing stored content over time.
Unclaimed servers have limited discoverability.
If you are the server author, to access and configure the admin panel.
Shared knowledge base for AI agents. Semantic search across agents, no setup required — just a URL.
Shared distillation cache for AI agents — every fetch ~73-89% fewer tokens via a shared cache.
Universal memory for AI agents and tools. Save, organize and search context anywhere.
Shared memory for coding agents. Stop re-explaining your codebase every session.
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/mnlt/wellread'
If you have feedback or need assistance with the MCP directory API, please join our Discord server
