declude

Before  "Great question! Before I answer, let me make sure I understand what
        you're asking. You want to know why the build is failing. Let me walk
        through my thinking, and then I'll give you the answer."

After   "I understand you want to know why the build is failing. Here's what
        I think is happening."

Before  "One thing I didn't touch, but you won't want to leave hanging, is the
        migration script. I'd be happy to walk you through it whenever you're
        ready!"

After   "One thing I didn't cover is the migration script. When you're ready,
        I can go through it with you."

Runs on an open-source model (Qwen2.5-14B-AWQ) on our own L4 GPUs, so your text never reaches a commercial AI provider. Prompt logging is disabled at the model server: text is processed in memory and discarded, never written to disk, a database, or logs.

Live: speak-english.tenken.co · Get an API key · Translate a document

Quick start

In Claude Code, install the plugin. It brings the skill and registers the hook:

/plugin marketplace add tenkenco/declaude
/plugin install declaude@tenken

Then run /declaude:setup. It safely migrates any manual hook and configures a dk_ key. Claude Code stores the key in secure plugin configuration. The hook rewrites replies as soon as a key exists. Set the hook_enabled option to false to stop it. The hook stays silent without a key, and it never blocks a session.

For the MCP server on its own, in Claude Code or any other MCP client:

claude mcp add --transport http declaude https://speak-english.tenken.co/mcp

No API key to paste: your client discovers OAuth, opens a browser sign-in, and holds the token. That is the whole setup.

Full usage, authentication and quota behaviour live in skills/declaude/SKILL.md — installable as an agent skill, so your agent can read it directly.

Related MCP server: Telnyx MCP Server

Architecture

client ── API key / OAuth / Clerk JWT ──> Cloud Run gateway ── VPC ──> internal LB ──> vLLM GPU MIG
                                              │                                    (Qwen2.5-14B, spot L4)
                                              ├─> Firestore (usage, paid flags, key hashes)
                                              └─> Stripe (payment link + signed webhooks)

• gateway/ — FastAPI. Every boundary is injectable, which is why the suite runs without network, GPU or cloud credentials.

• infra/ — Terraform, state in GCS, plan converges to zero diff. The GPU tier is private-IP spot instances behind an internal L7 load balancer.

• hook/ — Claude Code hook client. Fails open; never blocks a session.

• hooks/ — hooks.json, the plugin registration Claude Code reads on install.

• CI — lint, tests and coverage on every PR; security scans and a production smoke test daily.

• /healthz is reserved by Google Frontend on run.app — use /health.

• Model swaps are one Terraform variable; 32B needs 2×L4 (quota bump), 14B fits one L4.

• MIG template rollouts are deliberate (OPPORTUNISTIC): a proactive policy turned benign template edits into surprise 15-minute outages.

• Secrets live in Secret Manager only. API keys are stored as SHA-256 digests, so a database leak yields no usable credential.

• Prompt logging is disabled at the model server; request text is processed in memory and discarded.

• Developed test-first. Every production defect found while dogfooding became a regression test before its fix.

Development

cd gateway
uv sync --dev
uv run pytest -q          # 197 tests
uv run ruff check .

Credit

Grew out of gvzdv/claudish-to-english, the original local-Ollama hook. Licensed MIT; upstream notice in THIRD_PARTY_LICENSES.md.