Local security checks for what you build with AI, and a guard on every MCP tool call your agent makes.

SECOR^{(10.2026 – ∞Present)}

secor · 25 rules · Apache-2.0

Security for what you build with AI.

SECOR checks the APIs, MCP servers and websites you ship before they deploy, and guards every tool call your agent makes to the MCP servers it connects. Local and deterministic, and it never calls anything secure.

secor ship

local

1

2

3

4Highsecret in git history

5

6

7Medany-origin CORS

8

9

10

11

25rules

core5api9mcp5web6

Insights

Tests passing: 1,36858 files, 11 s
Rules: 25core · api · mcp · web
Planted leaks caught: 100/100secrets and personal data
False blocks: 0/250legitimate tool calls

From a checkout of the repository, then:

secor init                 # config and .secor/.gitignore
secor ship                 # check the project
secor gate --fail-on high  # for CI: exit 1 at that severity
secor dashboard            # local, read-only, on 127.0.0.1

Verified⁽²⁾

M8ven Trust Index: B, 89/100Verified Publisher · no concerning findings · 8 toolscode read 15 Oct 2026
1,368 tests passingUnit, fixture, CLI, Connect, Probe, MCP and dashboard suitesvitest · 16.10.2026
CI passingCI on main, GitHub Actions14.10.2026
Nothing leaves the machineNo telemetry, no LLM calls, redaction on every outputby design

Connect, live

secor connecthooks on · Claude Codereplayed from the hooks demo

askbrain_delete, left out of the scopetool-not-allowed
askbrain_remember with a customer's emailclass:personal_data
askbrain_capture with an API key in the noteclass:secret
passbrain_search with a normal query

Validation^{(9 repos)}

SECOR findings per stable rule in the first validation round, 9 public repositories, labelled by hand
Rule	Confirmed	False alarm
Vulnerable dependencies	83	0
Secrets in code or history	41	1
Unrestricted tool access	19	0
Security headers	12	0
Paid APIs without a rate limit	11	0

What I built

Local security tool for people who build with AI: it statically checks the APIs, websites and MCP servers you ship, vets the third-party MCP servers you connect to your agent, and guards every tool call that agent makes.
25 deterministic rules across Express, Next.js, FastAPI, Vite and the MCP SDKs, plus an opt-in prober that confirms a flaw against a staging target you own — and flatly refuses anything you can't prove you control.
No telemetry, no LLM in the engine, redaction on every output, and it never calls a system "secure". 1,368 tests, and it scans its own source clean.

TypeScript
AppSec
MCP
tree-sitter

Test count from `npx vitest run` in the SECOR repository on 16.10.2026. Catch and false-block rates from its Connect validation suite.

Validation round from DECISIONS.md: 9 public repositories, every finding labelled by hand.

Grade and score from the M8ven Trust Index.

Live figures as of 18.10.2026.

Next project

mail.google.com

लेRewriteShortenTranslateAsk

ClaudeGPTGeminiNanoyour key

CopyReplace

LekhakAI

A Chrome extension that sends one prompt to several LLMs from anywhere in the browser.

01.2025 —