Show HN: Algebras Harness – Localize Apps/Games from Claude Code

AI-powered localization workflow with automated QA and real-time fluency scoring via Algebras.

Works with Claude Code, Cursor, Codex, Windsurf, GitHub Copilot, or any agent that supports MCP tools.

AI agents: asked to localize a project with Algebras? Read AGENTS.md and follow it step by step. In short:

• Read the API key from ALGEBRAS_API_KEY (environment or .env).- Connect the MCP server https://platform.algebras.ai/api/mcp with header x-api-key: <key>.- Use the Project type: flow from plugin/PROJECT_TYPES.md (games, edtech, fintech, saas, or general).- Run the workflow in plugin/CLAUDE.md: onboard → glossary → translate → qa.

Open platform.algebras.ai/translation/agent , create an API key, and save it asALGEBRAS_API_KEY in your project's environment or.env . 2. Open your repo in Claude Code, Codex, Cursor, Windsurf, Copilot, or any coding agent, and paste the prompt from that page:

Localize this project with Algebras.
Follow the instructions in https://github.com/algebras-ai/algebras-agent
Use the API key from ALGEBRAS_API_KEY.
Project type: <games|edtech|fintech|saas|general>
Use the <type> flow from the agent instructions.

The agent connects to Algebras, builds your localization context, translates, runs QA, and leaves the result on a new git branch ( algebras/localize-<date> ) for you to review like any other change.

To install the agent permanently in your tool instead (plugin or MCP config), see Install.

Type	Use it for	The flow emphasizes
games	Games	UI, quests, item names, lore
edtech	Education products	Lessons, quizzes, course pages
fintech	Payments, banking, wallets, insurance	Money states, fees, disclaimers, statements
saas	Software products	App UI, emails, help center
general	Any other project	Detected from your files, per file

Each flow is defined in plugin/PROJECT_TYPES.md.

During onboarding the agent asks which fluency checks to run, and records the answer in project.json (add e.g. Fluency: context to the prompt to answer up front):

Setting	What is scored	Approx. cost per string of N words
none	Nothing	N
isolated	A no-glossary translation, scored on its own (context-free)	2N + 10
context	The written translation, scored with its neighbor strings and approved choices (context-aware)	N + 10
both	Both scores	2N + 20

Scores land in fluency_report.csv. They are a self-check of the agent's own translation, not an independent review.

Put any brief, style guide, character list, glossary, or screenshots in the project folder next to the translation files. Any file name is fine — the agent classifies files by their contents. What to include, and which facts each target language forces, is in plugin/CONTEXT_REQUIREMENTS.md. Missing context doesn't block translation: the agent asks at most five questions and writes the rest to client_questions.md for you to forward to the client.

| Glossary | Extracted from your code and UI strings, checked against established translations, and confirmed by you. Stored on the Algebras platform. | | Translation memory | Translations already in your repo are kept and reused as the reference for terminology and style. They aren't translated again. | | Context files | Specs, docs, lore and style guides found in your repo are classified by what they cover and read before glossary extraction and each translation batch. | | Context facts | Who speaks, to whom, formal or informal, what a placeholder stands for — recorded only from a cited source ( context_pack.json ,context_facts.jsonl ). Open points go toclient_questions.md , and the translation avoids committing to them. | | Screenshots | Screenshots already in your repo are linked to strings by column or file name and used as context. Automatic capture is coming soon. |

Note: Plugin commands require the Claude Code CLI (claude in your terminal). They are not available in the Claude Code web app or desktop app.

Run these commands in the Claude Code CLI from your project directory:

claude plugin marketplace add algebras-ai/algebras-agent
claude plugin install algebras-agent@algebras-agent

Then start an interactive session and run the setup skill — it opens your browser for the API key, copies workflow files into your project, and saves the key to .env:

claude
/algebras-agent:setup

The MCP server connects automatically once the plugin has your API key (declared in plugin.json). If you need to reconfigure the key later:

/plugin configure algebras-agent
/reload-plugins

Add the Algebras plugin marketplace, then install the plugin from Codex:

codex plugin marketplace add algebras-ai/algebras-agent
codex

Inside Codex, open the plugin browser:

/plugins

Install algebras-agent, then run the setup skill:

$setup

Restart Codex when prompted. You can verify the Algebras MCP server from:

/mcp

The plugin does not start the MCP server until setup saves your Algebras API key. If /mcp shows an unauthenticated algebras startup error from an earlier install, remove the old plugin install or delete the stale algebras MCP entry, run $setup, and restart Codex.

Run from your project root:

curl -fsSL https://platform.algebras.ai/install | bash

The script opens your browser for the API key, downloads the workflow files, and writes the MCP config automatically (.cursor/mcp.json, .windsurf/mcp.json, or ~/.codex/config.toml).

If your agent isn't auto-detected, pass --agent explicitly:

curl -fsSL https://platform.algebras.ai/install | bash -s -- --agent cursor
curl -fsSL https://platform.algebras.ai/install | bash -s -- --agent windsurf
curl -fsSL https://platform.algebras.ai/install | bash -s -- --agent codex

Once installed, open your project in your agent and send:

Translate this project.

Add a project type to pick a flow, e.g. Translate this project. Project type: saas. Or paste the full prompt from the Quick start — it works with or without the install step.

The agent onboards your project, builds a glossary, translates, and runs QA, asking you to confirm at each step. In a git repository it works on its own branch and commits the result at the end.

File	Purpose
CLAUDE.md	Workflow for Claude Code and Claude Desktop
AGENTS.md	Workflow for OpenAI Codex / Agents
.cursorrules	Workflow for Cursor
.windsurfrules	Workflow for Windsurf
.github/copilot-instructions.md	Workflow for GitHub Copilot
PROJECT_TYPES.md	The games ,edtech ,fintech ,saas andgeneral flows
CONTEXT_REQUIREMENTS.md	Client context the agent expects, the fact model, and which facts each target language forces
COMMON_MISTAKES.md	Error taxonomy from real-world LQA
skills/<phase>/SKILL.md	The phases: onboard ,glossary ,translate ,qa , plus optionalglossary-dedupe

AGENTS.md at the repository root is the entry point for agents that follow this repo's URL without installing anything.

Automate translation in any repo's CI pipeline — no manual setup required.

Add these secrets in your repo: Settings → Secrets and variables → Actions Secret Value ALGEBRAS_API_KEY From platform.algebras.ai/api-keysANTHROPIC_API_KEY From console.anthropic.com 2. Copy .github/workflows/translate.example.yml from this repo into your project's.github/workflows/ and adjust thepaths trigger to match your locale files. 3. Push to main (or trigger viaActions → Translate → Run workflow ).

On each run the action:

1. Downloads the latest workflow prompt files from this repo
2. Checks project.json — exits early if all strings are already translated (skip-if-complete: true )
3. Writes a temporary MCP config pointing to the Algebras platform
4. Installs the Claude CLI and runs it non-interactively against your project
5. The agent runs the full workflow: Onboard → Glossary → Translate → QA (skips phases that are already done)
6. Detects changed files via git diff ; commits and pushes (or opens a PR)

Input	Default	Description
algebras-api-key	—	Required. Algebras API key
llm-provider	anthropic	anthropic orgemini
anthropic-api-key	—	Required if provider is anthropic
gemini-api-key	—	Required if provider is gemini
algebras-platform-url	https://platform.algebras.ai	Override for self-hosted
commit-changes	true	Commit translated files
create-pr	false	Open a PR instead of pushing directly
skip-if-complete	true	Exit early if nothing needs translating
project-type	general	Flow from PROJECT_TYPES.md :games ,edtech ,fintech ,saas , orgeneral
fluency	isolated	Fluency checks: none ,isolated (context-free),context (context-aware), orboth
min-fluency-score	(empty)	Fail if any string scores below this (1–10)

- uses: algebras-ai/algebras-agent/.github/actions/algebras-translate@main
  with:
    algebras-api-key: ${{ secrets.ALGEBRAS_API_KEY }}
    anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}
    commit-changes: "true"
    create-pr: "false"

• Python 3.10+ (for QA tools)
• An Algebras account

source & further reading

github.com — original article I built Giverny – a native terminal for Cluade Code and Codex Show HN: Halis Code – Agent Coding Show HN: OpenRMM – open-source RMM where the AI proposes and a human approves