Missions, not sessions

Fire the mission. Come back to done.

Fire work at an agent under an envelope you wrote. Detach. It interrupts you only when the envelope says so. Or run it headless over a task list, one line at a time, each item in a clean context.

Run a recipe View on GitHub

$ curl -fsSL https://contenox.com/install.sh | sh

Installs contenox. Run contenox init to scaffold a workspace.

any model one binary no account SQLite your rules in a file you can read

The terminal UI

One session. Chat, plan, shell.

contenox new opens a persistent terminal session where conversation, plans, and shell commands share one context — and MCP tool connections stay alive across turns instead of reconnecting on every call.

contenox new in the terminal: a natural-language task, contenox reads the repo and answers in the same session

Chat, plan, and run in one place

Talk through the problem, sketch the plan, execute the commands — without switching windows or losing context between modes.

MCP stays warm

Tool servers connect once and stay connected across turns, so stateful tools keep their state and answers come back fast.

Approvals that stay out of the way

Risky steps pause for a quick yes or no right in the session — sensible defaults, no ceremony. A compound line of allowlisted commands just runs; the envelope interrupts only where it must.

Explore the contenox CLI →

Core capabilities

Close the terminal. It can’t wreck anything while you’re gone.

Tool allowlists, command policy, budgets, approval gates — what the agent may do unattended is a file you wrote, not a guess. An unanswered approval checkpoints the run and frees the seat; answer it from any terminal, even days later, and the run resumes exactly once.

The chain is the contract

Prompts, model routing, and every rule are visible keys in a versioned file — reviewable, diffable, repeatable. One harness behind the CLI, the contenox TUI, and any ACP editor.

It reads Go like the compiler

Describe, definition, references, implementations, symbols, diagnostics — six queries answered in-process by the real type-checker, warm in microseconds. No language server to install or babysit.

Ask the repo, get cited lines

contenox index builds a hybrid full-text and embedding index in local SQLite; contenox search answers with file:line citations and marks a hit stale the moment its file changes.

Providers are config

Route inference to local Ollama or vLLM, or to OpenAI, Anthropic, Gemini, Bedrock, or Vertex — without touching the workflow.

Local by default

Sessions, config, run logs, and state live in SQLite on your machine. No hosted Contenox service, no account — telemetry is opt-in and off by default.

Terminal demo: a destructive rm command stops at a human approval gate before it runs

Fire one, detach, come back to done. The envelope bounds what the agent may do while you’re gone; the chain is the contract for what it does. Same sessions in your terminal and your editor, any model and any MCP in combination. Built by its first customer — who is hard to convince.

The old way	contenox
A hidden prompt	A file you edit
Guessing the blast radius	An envelope you wrote
Watching it work	Detached missions
One vendor	Any model, any MCP, together

What you can author

Repeatable work, with policy in the file.

→Release evidence pack: feed in git history, PRs, ticket notes, and CI output; the chain drafts the changelog, risk notes, and deployment checklist, then stops before publishing.
→API-to-workflow wrapper: register a narrow OpenAPI subset, inject hidden workspace or environment arguments, and require approval for mutating calls.
→Repo maintenance chain: inspect an issue, plan the patch, touch only approved files, run tests, and prepare a review packet instead of silently changing the project.
→Browser QA evidence: Playwright captures the live app state, the chain writes documentation or screenshots, and the policy decides what needs human approval.

Run it where work happens

The same chain can run in your editor, terminal, or automation.

Contenox speaks the Agent Client Protocol over stdio. Use Zed, JetBrains, AionUi, or OpenClaw when an editor UI helps; keep the workflow itself in the chain file, with tool calls and HITL approvals routed through the client.

// ~/.config/zed/settings.json
{
  "agent_servers": {
    "Contenox": {
      "type": "custom",
      "command": "contenox",
      "args": ["acp"]
    }
  }
}

Read the Zed guide →Read the JetBrains guide →Read the AionUi guide →Read the OpenClaw guide →

Built on open standards

Go the implementation SQLite local state Ollama local models vLLM local models

Local models connect through Ollama or vLLM over their local APIs; hosted providers over their public APIs. No bundled inference runtime, no GPU drivers to ship.

Get started

Fire your first mission.