About codeitall
codeitall is a hosted MCP server for AI coding agents working with Rust crate APIs. Its version-pinned responses report observed runtime behavior alongside API status information, using deterministic retrieval rather than generative answers.
Key features
- 01Answer open-ended Rust API questions
- 02Search by function signature shape
- 03Look up observed function behavior
- 04Compare crate implementations
- 05Find current alternatives to deprecated calls
Use cases
- 01Check a Rust function's behavior on inputs
- 02Compare crate behavior for a task
- 03Find a replacement for a deprecated API
MCP support
Exposes five Rust API query tools to MCP clients over SSE/HTTP; stdio is available for self-hosters. The same tools are also accessible through HTTP REST.
Configuration
Remote servers need no local install. Pick your client, then copy the config or install in one click.
https://api.codeitall.dev/mcpclaude mcp add --transport http dev-codeitall-codeitall https://api.codeitall.dev/mcp{"mcpServers":{"dev-codeitall-codeitall":{"type":"http","url":"https://api.codeitall.dev/mcp"}}}Open in Cursor{"servers":{"dev-codeitall-codeitall":{"type":"http","url":"https://api.codeitall.dev/mcp"}}}Open in VS Code{"mcpServers":{"dev-codeitall-codeitall":{"type":"http","url":"https://api.codeitall.dev/mcp"}}}{"mcpServers":{"dev-codeitall-codeitall":{"serverUrl":"https://api.codeitall.dev/mcp"}}}Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.
Tools
Details
From the project READMEcodeitall — Rust API behavioral oracle
Version-pinned behavioral oracle for Rust package APIs, built for AI coding
agents. Five deterministic tools backed by a Postgres substrate of
66,549,370 pattern_records and 107 successful probe runs across
7 task families (datetime, duration, base64, JSON strict-vs-permissive,
JWT verify, AEAD encrypt/decrypt, constant-time compare).
Hosted, free, no signup: https://api.codeitall.dev
This repository is codeitall's public surface — the API reference, OpenAPI spec, the "For AI Agents" page, and the per-family pages. The ingest/indexer and server source are not redistributed; you consume codeitall by querying the hosted API, not by running it yourself.
What it answers
The differentiator from a docs RAG: the substrate captures runtime behavior (what a function actually does on a canonical input set, observed empirically), not just documentation (what it claims to do). It targets the documentation-vs-runtime gap — yank status, RustSec advisories, deprecation renames, and cross-implementation behavior tables.
Tools
| Tool | Use it for |
|---|---|
answer_api_question | Open-ended Rust API questions; routes deterministically (rule-based, no LLM) and falls back to bge-m3 cosine retrieval. |
signature_search | You know the signature shape (fn(&str) -> Result<Vec<u8>, _>); get ranked candidate functions + crate, downloads, yank status. |
behavior_lookup | A specific (crate, fn_name); get the probe-observation table verbatim — each row (input, outcome=ok|err|panic, value_or_error). |
compare_implementations | A task category (e.g. "base64 crates"); get the side-by-side cross-implementation behavior table. |
find_modern_equivalent | A suspected-deprecated call; get the current canonical alternative + yank/advisory rows. |
Determinism contract
No generative LLM sits between a request and its response. Every response is SQL + deterministic templating; the only model in the hot path is bge-m3, a deterministic text→vector embedder used for the cosine fallback (same input → same vector → same hits). This is a regression-pinned invariant.
Quickstart
# Mint a free anonymous key (no email, no signup):
curl -sX POST https://api.codeitall.dev/v1/keys \
-H 'content-type: application/json' -d '{}'
# Ask a question (HTTP REST):
curl -sX POST https://api.codeitall.dev/v1/answer_api_question \
-H 'content-type: application/json' \
-d '{"question":"how do I parse a duration string in rust"}'
GET aliases exist for every tool (browse-only clients); signature_search
also accepts a shape_b64 (base64url) parameter so URL re-encoding can't
mangle type-signature characters.
Access modes
- MCP — SSE/HTTP at
https://api.codeitall.dev/mcp(Bearer key), or stdio for self-hosters. - HTTP REST —
POST/GET /v1/<tool>. - OpenAPI 3.1 —
openapi.json(live:https://api.codeitall.dev/openapi.json).
Docs in this repo
for-ai-agents.html— positioning, tool list, key snippet (live:https://api.codeitall.dev/for-ai-agents.html).api-reference.html/api-reference.md— generated tool reference + scope/coverage.agents/— one page per probe family.
Cost & limits
Free. Per-IP key issuance cap (10/day) and per-key quota (10,000/day). No paid tier, no email capture, no signup.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.