Contenox
An agent harness for policy-bound, vendor-agnostic AI-assisted software engineering.
Contenox runs authored agent workflows as a local process. The policy that gates
each tool call is a file the operator wrote, enforced by the harness before
anything runs. Backends are configuration; sessions, approvals and per-step
records stay on disk.
The harness is the product. One binary serves the CLI and a local cockpit,
editors attach over ACP, and the model named in a request is the model that
serves it — or the run stops with the candidates it could have used.
What Contenox is not: it is not an autonomous coding employee, not a hosted
autopilot, and not a vendor's harness. It is a local runtime for AI-assisted
work that keeps the operator accountable for what runs.
Docs: contenox.com
For Work That Touches Real Systems
Contenox is purpose-built for high-consequence engineering environments: production repositories, internal APIs, infrastructure scripts, operational dashboards, release pipelines, and systems of record.
In these environments, an "AI agent" cannot mean "give a model broad credentials and hope for the best." It requires a strict runtime boundary:
- Explicit tools & policies to govern actions.
- Local state to preserve privacy.
- Human-in-the-loop (HITL) approval for execution.
- Reviewable evidence for auditing.
The core issue is not that agents can act; it is agents acting outside a boundary the operator authors. The model can reason, inspect, and propose—but the Chain decides what it may touch, and the operator decides what it may change.
Install
curl -fsSL https://contenox.com/install.sh | sh
Inspect Before Installing
If you prefer to audit the installation script first:
curl -fsSLO https://contenox.com/install.sh
less install.sh
sh install.sh
Pre-built release downloads and source builds are also available on the releases page.
Quick Start
contenox setup # Choose a provider/model for this machine
contenox "say hello world in python" # Query directly from the CLI
contenox chat -e # Open $EDITOR to compose a rich prompt
Manage past contexts effortlessly using contenox session list and contenox session switch <name>.
Smart Dual-Model Routing
Inline autocomplete runs on a dedicated model separate from the main chat. This ensures editor ghost text stays entirely local and ultra-low latency, while complex chat queries can leverage larger frontier models:
# Rich chat routed to a hosted model
contenox config set default-provider openai
contenox config set default-model gpt-5-mini
# Ghost text routed to a small, fast model
contenox config set default-autocomplete-provider ollama
contenox config set default-autocomplete-model qwen2.5-coder:7b
Core Use Cases
Contenox excels when workflows are specific, repeatable, and require explicit guardrails:
- Reviewing Diffs: Run tests, summarize architectural risks, and gate destructive operations behind manual approvals.
- Drafting Release Evidence: Automatically aggregate git logs, PRs, issue tickets, and CI outputs into clean changelogs, deployment checklists, and reviewer packets.
- Wrapping Internal APIs: Safely expose subsets of OpenAPI specs while masking sensitive tenant/environment arguments and requiring authorization for mutating calls.
- Automating Repo Chores: Ingest an issue tracking item, generate a patch, run local validation checks, and draft the PR description.
- Inspecting Live Operations: Query diagnostic dashboards, shell scripts, or Model Context Protocol (MCP) tools via tightly scoped policies rather than broad, persistent credentials.
- Edge Autocomplete: Offload editor suggestions to local workstation models while maintaining high-powered reasoning in chat.
The exact same Chain runs seamlessly across the Terminal, Zed, JetBrains, AionUi, or any standard ACP client.
Security & Runtime Protection
| Identified Risk |
Contenox Mitigation |
| Fleeting Agent History |
Chains are declarative files: easily reviewable, version-controlled, and repeatable. |
| Unbounded Agent Access |
Enforced via strict tool allowlists, localized command policies, and tightly scoped API definitions. |
| Post-Damage Review |
Destructive actions are systematically blocked by human-in-the-loop approval gates prior to execution. |
| Leaky Internal APIs |
Curated OpenAPI subsets encapsulate hidden environment variables, auth tokens, and tenant arguments. |
| Vendor Lock-in |
Provider and model routing live in configuration files, entirely decoupled from application logic. |
| Frontier Model Budget Burn |
Routine tasks and linting checks are automatically routed to local or private-network infrastructure. |
| Exfiltration of Team Data |
All interactive sessions, states, configs, and runtime logs remain completely local. |
Architectural Fit
Contenox serves as the local agent runtime layer running between your interface and your infrastructure.
| Ecosystem |
The Contenox Paradigm |
| IDE Copilots |
Editor assistance is only one piece of the puzzle. Workflows must execute uniformly across the terminal, IDEs, and independent headless scripts. |
| CLI Coding Agents |
A single coding loop is not a structured runtime. Contenox adds multi-session persistence, strict tool authorization, explicit model routing, and human gates. |
| LangChain / Frameworks |
Software development libraries are not end-user products. Contenox provides an out-of-the-box executable runtime tailored for engineers and teams. |
| Dify / n8n / Web Tools |
Workflows touching local source code and specialized internal infrastructure should never depend on a third-party SaaS control plane. |
| Ollama Wrappers |
A model provider is not a workflow boundary. Contenox introduces native Chains, secure tool definitions, and dynamic routing across hybrid infrastructure. |
Connect Your Stack
Any asset accessible via an OpenAPI spec, a shell command, or an MCP server can be transformed into a secure tool inside a Contenox Chain:
# Connect any Model Context Protocol (MCP) server
contenox mcp add notion https://mcp.notion.com/mcp --auth-type oauth
# Wrap an internal HTTP API using its OpenAPI specification
contenox tools add erp_billing \
--url https://erp.internal.example.com \
--spec ./billing-subset.yaml
# Bind the local shell under a strictly defined Chain policy
contenox --shell "check Proxmox and flag anything red"
Editor Integration
Contenox natively communicates via the Agent Client Protocol (ACP) over standard I/O.
Zed Integration
Add the following snippet to your ~/.config/zed/settings.json:
{
"agent_servers": {
"Contenox": {
"type": "custom",
"command": "contenox",
"args": ["acp"]
}
}
}
Tool invocations render dynamically as interactive UI cards showing the underlying paths, human-in-the-loop prompts hook directly into the editor's native permissions, and full session history replays automatically upon reopening projects.
Other Environments
- Web Interface: Running
contenox serve mounts your active sessions, chains, and backends locally inside Beam, the bundled web UI.
Step-by-step guides: Zed | JetBrains | AionUi.
Managing Backends
Routing rules are treated as system configuration, not application logic. Add hosted, local, or private network backends seamlessly:
# Private infrastructure & local inference
contenox backend add ollama --type ollama
contenox backend add myvllm --type vllm --url http://gpu-host:8000
# Commercial cloud providers
contenox backend add openai --type openai --api-key-env OPENAI_API_KEY
contenox backend add anthropic --type anthropic --api-key-env ANTHROPIC_API_KEY
contenox backend add gemini --type gemini --api-key-env GEMINI_API_KEY
# Set global operational routing defaults
contenox config set default-model qwen2.5:7b
contenox config set default-provider ollama
Building From Source
Prerequisites
# Clone the repository
git clone https://github.com/contenox/harness
cd runtime
# Build the main core CLI
make build-contenox
Core Dependencies
The contenox core CLI is written entirely in pure Go. Go dependencies are maintained standardly in go.mod.
Questions? Reach out at hello@contenox.com