Self-hosted
Your agent on rails, on your hardware, with nothing leaving the building
Most agent tooling assumes a connection to someone else's server: for the model, for the orchestration, for the license check. If your code cannot leave the network, that rules out the tooling before you have evaluated it. CodyCody is built the other way round. The engine is a binary you run. It activates from a signed license key without contacting us, and it has nothing to contact us for afterwards.
What you get is the same engine and the same workflows as everyone else — deterministic step order, shell steps under an exit-code contract with output hashed to file, human approval gates — running entirely inside your perimeter, against whatever model endpoint your perimeter allows.
Security and compliance review increasingly asks for something more specific than data residency: AI agent governance — who approved what, what actually ran, and proof of it, kept somewhere your team controls rather than a vendor's. Self-hosting is the precondition for that. The engine's own instance log is what actually delivers it.
The facts
- ✓Offline activation with a signed license key; the engine never phones home.
- ✓Self-hosted on Windows, macOS and Linux; the engine bundles its own runtime.
- ✓Enterprise adds air-gapped operation, SSO, audit-log export and a site license.
- ✓Built in the Netherlands, sold by an EU company, with EU VAT handled at checkout.
- ✓Every run's instance log records step results, gate waits, review verdicts and forced approvals — the audit trail a regulated team needs, held on the instance you control.
The problem
Where it goes wrong without rails
Security review asks what data the tool sends home, and the honest answer for most agent tools is "your repository".
A license server that must be reachable is an outage waiting to happen, and a non-starter on an isolated network.
Regulated teams need an audit trail of what the agent did, and a workflow that runs in someone else's cloud cannot give them one they control.
"Where does the audit trail live" is a question a cloud-hosted agent tool answers with "in our systems" — the wrong answer when the requirement is that it doesn't leave yours.
How it works
What the engine does about it
- 1
Download a build for your platform
The engine ships as precompiled bytecode with the exact Node.js runtime it was built for — one build each for Windows x64, macOS Apple Silicon and Linux x64. There is nothing to install around it.
- 2
Activate offline
Paste your license key into the engine's configuration. The key is signed; the engine verifies the signature locally and refuses to start without one. No activation call, no heartbeat.
- 3
Point your agent at localhost
The MCP server listens on your machine. Claude Code, Codex or any MCP client connects over streamable HTTP to a local address with an agent token you issued — the model endpoint is whatever your agent is already configured to use.
- 4
Run workflows that never leave
Every step, every shell result and every approval decision happens on your hardware. Captured output lands in files on the host; the findings a review workflow produces are written into your repository, not to a service.
- 5
Every step lands in an audit trail you hold
ccw_instance_logs returns the full record of a run — every step started and completed, every gate wait and its answer, every forced approval when a review loop hit its retry limit, and the reason given for any cancellation or rewind. It's read from your own instance. Enterprise adds bulk export for retention and review.
Questions
Asked before buying
- Does the engine ever contact CodyCody?
- No. Activation is a local signature check on your license key, and nothing in the engine reports back afterwards. Updates are downloads you choose to fetch from your account page.
- Can it run with no internet connection at all?
- The engine can. Whether the agent can depends on where its model runs: with a model endpoint inside the network, the whole loop is air-gapped. Enterprise is the tier built for that deployment.
- What about the visual workflow designer?
- It is served by the engine itself, on a local port, so it is self-hosted too. Workflows and skills you author stay on your instance.
- Does this cover AI agent governance, or just where the software runs?
- Both are the same requirement seen from different angles. Governance needs a record of what an agent did and why, and a way to stop it before something irreversible; self-hosting decides where that record lives. The engine supplies the record — the instance log — and the stop — approval gates. Self-hosting keeps both inside your perimeter instead of a vendor's.
- What's actually in the audit trail?
- ccw_instance_logs returns every step's start and result, every gate wait and who answered it, every forced approval from a review loop that hit its retry limit, and the reason given for any cancellation or rewind. It's read from your own instance — nothing is summarized or discarded to save space on our side, because there is no side to save it on.
Start with
The packages that fit
- Security Review →
A source-level OWASP Top 10 assessment and a STRIDE threat model with trust boundaries, built by your coding agent, step by step. Included with Pro.
- Code Quality →
Run a full ISO/IEC 5055 and Clean Code assessment of any codebase with your coding agent, every check its own verified step. Included with every license.
Every plan and what it includes is on the pricing page.