# koda [Health: Active]

**Category:** 🧠 Knowledge & Memory  
**Repository:** https://github.com/aabxtract/koda  
**GitHub Stars:** 0  
**npm Downloads (last month):** 210  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/koda

## Description
Verification and memory layer for coding-agent workflows.

## Tools
Capabilities this server exposes over MCP:

- **koda_verify** — Run focused verification. The agent **supplies its own impact analysis** (`flows`, `endpoints`, `target`, `commit`, `max_flows`) — it already read your code, so no LLM cost and no guessing. Groq is only a fallback for non-MCP runs
- **koda_report** — Latest full JSON report — verdicts, coverage, evidence paths
- **koda_memory** — Project memory — recurring failures, fix confidence, time-to-fix
- **koda_setup_cicd** — Generate a validated GitHub Actions workflow

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "koda": {
    "command": "npx",
    "args": ["-y","koda-verify"]
  }
}
```

## Documentation

## What koda MCP server does

The koda MCP server provides a verification and memory layer for coding-agent workflows. It does not edit application code. Instead, it coordinates checks for a selected commit or change, writes machine-readable and human-readable reports, and records recurring failures and their resolution history.

Verification is scoped to the supplied impact analysis. An agent can identify relevant browser flows and endpoints after reading the code, then pass that information to Koda rather than asking another model to infer the impact. The resulting report includes verdicts, coverage information, evidence paths, risk data, and any action required.

## How it works

Koda runs as a stdio MCP process started through the `koda-verify` package. The normal workflow is for an agent to edit and commit code, while Koda's hooks run verification in the background. The agent can then call the MCP tools to run focused checks or inspect the latest results.

Checks are selected according to changed files. API, configuration, and test-only changes can use endpoint checks and integration tests. UI-related changes can trigger headless Kane browser flows scoped to the affected pages or components. Endpoint verification uses GET, HEAD, and OPTIONS by default. Supported test runners include Jest, Vitest, Mocha, and node:test when tests match the changed files.

Each run produces JSON and Markdown files under `.koda/reports/`; browser session evidence is retained under `.koda/evidence/`. Project memory tracks stable failure keys, repeated patterns, passing fixes, and time to resolution. A skipped check, such as one affected by an offline application, is not treated as a failure or added to memory.

## Setup and configuration

Install the package globally with `npm install -g koda-verify`, initialize a project with `koda init`, and use `koda doctor` to check Git, configuration, the target URL, directories, and Kane availability. The MCP process can be registered with a client using `claude mcp add koda -- npx koda-verify mcp`.

The koda MCP server uses the project's `.koda` directory for configuration, reports, evidence, and memory. Its target can be saved with `koda target <url>`, supplied for one run with `--target`, or set through `KODA_TARGET`. Target precedence is command-line flag, environment variable, project configuration, then automatic detection of common local development ports.

Browser verification is optional. Kane can be installed separately and authenticated with the user's own account. Without Kane, endpoint checks, integration tests, reports, and memory remain available while browser checks are marked `SKIPPED`.

## Tools and capabilities

- `koda_verify` runs focused verification using agent-provided `flows`, `endpoints`, `target`, `commit`, and `max_flows` values.
- `koda_report` returns the latest full JSON report, including verdicts, coverage, and evidence paths.
- `koda_memory` exposes recurring failures, fix confidence, and time-to-fix information.
- `koda_setup_cicd` creates a validated GitHub Actions workflow.

The generated workflow can install dependencies, build when applicable, install and authenticate Kane through GitHub secrets, start the application, wait for its target, and run verification for the current commit. The template refuses to overwrite an existing workflow unless forced.

## Limitations and notes

The koda MCP server supplies verification and reporting rather than code generation or code changes. Browser flows depend on Kane and the user's separate Kane authentication; no credentials are bundled. HTTP checks are read-only unless mutating methods are explicitly enabled and endpoints are declared by the agent.

Existing Git hooks are backed up before Koda appends its hooks, and the setup is designed to be idempotent. If an application is unavailable, Koda reports checks as skipped rather than failed. Groq is only a fallback for non-MCP impact analysis; MCP verification relies on the agent's supplied analysis.

_Full upstream README: https://allmcps.com/mcp/koda/readme_