Ship with better understanding of dev docs using Context7

[Juice10]

Context7 seems a pretty powerful MCP server that includes docs for most dev projects. This is something that is helpful for most developers, so it would be great if Kilo Code shipped with it by default.

Of course we should play with it first to see if it lives up to its promise. EDIT: Looks like a big improvement

From https://www.reddit.com/r/ChatGPTCoding/comments/1k6uyjp/i_just_found_out_about_context7_mcp_server_and/?share_id=KY3QTk-umnuf4DFwizogh&utm_content=1&utm_medium=ios_app&utm_name=ioscss&utm_source=share&utm_term=1

Ticket: #655

[Juice10]

I've experienced a 20% cost reduction when using context7 and the following custom instructions for all models:

When are attempting a change, please look up the project's dependencies and use the context7 mcp to read the current api documentation for the relevant libraries to make sure you are using the these libraries and their latest best practices correctly

[druellan]

I used Context7 for a while, but I ultimately replaced it with a search + fetch MCP. While Context7 is free and usually returns a more concise and economic-formatted information, it is as good as the API database it has. If you are working with an API that is not in base, you are out of luck.

My feeling is that Kilo can benefit of a built-in web search and a webfetch mechanism. Sadly, most of the web search APIs are not free or at least they require a login. Copilot, for example, opted to use Tavily or Bing for that. I'm using Tavily right now, and while I can't say their search is good or not, the MCP offers everything you need to consume information from the web, that I think it is the point.

Said that, for out-of-the-box-you-dont-need-to-do-anything solution, Context7 is solid, it is free, don't ask questions, and it is built for AI context.

[Triquetra]

I haven't tried a websearch MCP, but I'm skeptical of the value. I think the problem with a more general web search is the volume of problematic data. Much of the data returned by search engines is:

• useless: AI generated fluff or otherwise non-substantive articles that have the right keywords
• outdated: maybe relevant 3 years ago, but not anymore, even though it is popular
• misleading: it looks like it's related, but it's really about something else
• wrong: it is exactly on point, but it gives broken code, violates best practices, or provides incorrect information

So, while the web search will provide the AI with a volume of additional data (consuming your context and tokens), the question remains whether the AI will be able to sort out the above problems with general internet data correctly.

Using context7 or deepwiki is essentially a hyperfocused internet search that has been intentionally limited to quality, relevant information.

context7 provides the AI very clear information about current, official documentation for libraries. The one downside to context7 is that information retrieval is siloed as far as I can tell -- you can only get information about one library at a time, and unless that library's official documentation discusses interaction with another library, the AI will need to make an additional request for another library you are working with and infer from official documentation how the two may interact.

IMO, that context7 weekness is where deepwiki shines. It seems to provide more than official documentation. It provides usage examples for the libraries that you are using drawn from many other open source projects that use the same libraries.

So, I'm increasingly using the two in this way:

• context7: official documentation and syntax reference
• deepwiki: usage examples and debugging verification

[Pokeylooted]

You can always use stagehand mcp, allows it to control a full browser using playwright and lookup docs that way. Duckduckgo search is also free iirc

[Triquetra]

@Juice10
I've been using this Anti-Hallucination rule, which has helped. It doesn't always use a tool, but it usually does when it should. Occasionally, I feel like I still have to prompt it to "use context7", but most of the time that I do, it doesn't change the outcome.

# Anti-Hallucination Rule for Kilo Code

## Rule Statement
**ACCURACY VERIFICATION REQUIRED**: Before providing any factual information, code examples, API references, library documentation, or technical specifications, you must verify the accuracy of your response. First, assess confidence in your training data knowledge. If you have high confidence in well-established, stable information from your training, you may provide it directly. If uncertain about accuracy, currency, or specifics, use appropriate available tools (such as context7, deepwiki, web search, or documentation access) to verify before responding. If neither training data nor tools can confirm the information, state your uncertainty and recommend independent verification.

## Implementation Guidelines
- **Training data first**: Rely on training data for well-established, stable technical concepts and widely-used APIs/libraries
- **Tool verification when uncertain**: Use context7, deepwiki, web search, or other available tools when:
  - Information may have changed recently
  - Dealing with specific version details, recent updates, or emerging technologies
  - Uncertain about syntax, method signatures, or current best practices
  - Asked about specific error messages, configurations, or recent changes
- When using tools for verification, cite the source used
- If neither training data confidence nor tool verification is available, acknowledge uncertainty explicitly
- Direct users to official documentation for definitive answers when confidence is low

## Priority
This rule takes precedence over providing quick answers - accuracy over speed.

[gnewtonn]

Nice! Do you keep it in Global or by Project rules?

[Triquetra]

Global rules

[Triquetra]

I don't currently use web search, but I use both context7 and deepwiki, and I find them both valuable.

[adamhill]

I tried C7 heavily while working on my MCP to control Rebelle through its new websocket interface.

C7 was the difference between a terrible MCP server with basic websocket functionality and a nice FastMCP implementation with a robust websockets implementation using the Websockets library.

The main annoyance was to remember to add some admonition to the prompt to use Context 7, whenever it need to look up docs.

Definitely a useful and affective MCP.

NB. I also think Frontmatter (#1269) could narrow the focus for language specifics, much like C7 does for framework specifics. Imagine sticking an @http:// URL in there and running it through C7 automatically some other signally mechansim to trigger retrieval.