AKAMAN MCP Interface

The Model Context Protocol is a structured way for an AI application to talk to a server. MCP gives the application a repeatable contract:

1. connect to a server,
2. discover the tools the server exposes,
3. inspect each tool’s argument schema,
4. call a tool with structured arguments,
5. receive a structured result.

Coding agent / AI application
        │
        │ MCP client
        ▼
    MCP server
     (AKAMAN)
        │
        ▼
local system documentation

For AKAMAN the interface is intentionally tiny. It exposes exactly one MCP tool:

man(query, section?, source?)

query

Required. The thing being looked up — a command, an option, a named section, or a header symbol.

section

Optional. A numeric manual section (for example 2 for system calls) or a named section heading (for example exit status).

source

Optional. Explicitly selects man, doc, or headers. A selected source is never silently substituted.

The three sources

• man — the native installed manual pages. AKAMAN delegates discovery and rendering to the host’s own man rather than building a second index.
• doc — installed documentation under /usr/share/doc: plain text, Markdown, READMEs, and HTML files.
• headers — declaration blocks from installed C/C++ system headers.

Progressive disclosure

This is one of AKAMAN’s strongest ideas. A bare query such as akaman grep can return the synopsis plus a compact map of the real section names in the page. The agent can then make a second, targeted request — akaman "grep exit status" or akaman "grep -r" — and receive only that part. The agent discovers what is available first, then asks for exactly the detail it needs. That is far better than dumping a whole manual into context “just in case”.

Focused answers are budget-capped at roughly 400 estimated tokens, so an agent can call AKAMAN repeatedly without letting reference material dominate its context.