Last updated · Santosh Kumar Radha

A coding agent that keeps your code in your own infrastructure

What leaves your machine with CodeAF, how to keep code off third-party model APIs with local or self-hosted models, and what air-gapped use needs.

CodeAF, the open-source (Apache 2.0) coding agent by AgentField, runs on your machine and sends your prompts and code only to the model provider you connect; point it at a local or self-hosted model and those model calls stay inside your network (FAQ). Two other streams leave by default, and both carry numbers, never code: anonymous usage counts and Model Pool scores. This guide lists each one, how to turn it off, and what a fully air-gapped setup does and does not have documented.

At a glance: what leaves the machine

StreamWhat it carriesNever carriesTurn it off
Model callsYour prompts and codeUse a local or in-network endpoint
Usage countsVersion, OS, mode, session counts, errors, total tokens as counts and bands, under a random per-install idPrompts, code, file names, paths, repo names, keys, email, IP address, machine name, model namesCODEAF_TELEMETRY=off, DO_NOT_TRACK=1 or the telemetry switch in /settings
Model PoolMetric, seat role, model, score, which model judged, kind of run, size bucket, day, under a random per-install idCode, prompts, file names, paths, anything about youmodel_pool = off, or read to fetch the index and send nothing
Connected accountsOnly what you connect with /connectDo not connect them; keys stay in your profile

Sources: FAQ, Model Pool, commands, connections, and the README's telemetry section.

Step 1: choose where model calls go

The model provider sees what the agent sends it. To keep code away from OpenAI, Anthropic or any other API, connect a model you run: Ollama needs no key, and any OpenAI-compatible endpoint works, such as llama.cpp or LM Studio, "so model calls stay on your machine" (FAQ). The same holds for an inference server elsewhere in your network that speaks the OpenAI-compatible API. /crew models open keeps every crew seat on open-weight models (README), and open-weight coding models covers which models to expect results from.

Check that nothing else is connected: codeaf connect lists every model service and whether it is connected, and codeaf disconnect <service> removes one (commands).

Step 2: turn off the numbers too

bash
export CODEAF_TELEMETRY=off

That stops the anonymous usage counts; DO_NOT_TRACK=1 or the telemetry switch in /settings does the same. The README adds that any one of them stops every stream, the Model Pool's rows included, and that the full field list is in docs/TELEMETRY.md.

The Model Pool has its own setting, and it is off by default: no pool judging, recommendations or sharing. on uses and contributes text-free scores, read fetches the shared index and sends nothing, and off turns the pool and its judge off (Model Pool). The judge matters for privacy too: after a task lands, a model outside the task's crew scores the work, and that is one more call to your provider. codeaf pool status shows this install's mode and how many rows wait to be sent, and ~/.codeaf/pool/outbox.jsonl holds the rows themselves so you can read them before they go.

Step 3: keep the work on a machine you control

Code, conversations and tasks live on the machine that runs CodeAF. For a team, put that machine inside your network and reach it over your own ssh with codeaf chat --host devbox, with no daemon and no port to open (remote machines, SSH guide). Accounts such as Linear or Slack reach the agent only if you connect them; their keys and tokens stay in your profile on that machine, and tasks cannot add new connections (connections).

Air-gapped environments: what is and is not documented

Documented: CodeAF is one binary with nothing else to install; the releases page holds the assets and checksums for each platform, and you can build it yourself with make build (README). Local models need no key. senior-dev's cached model capacity works offline, and it assumes a 16,384-token context only when no model catalog can be read (senior-dev).

Not documented: an official offline installer or an air-gapped install procedure. The install script downloads from the network, codeaf update and codeaf models --refresh reach out, and we have not published a tested air-gapped run. If you need one, the pieces above are where to start, and asking on the AgentField Discord is the honest next step. For approvals and records inside a regulated network, see approvals for regulated teams.

When something else fits better

Tabby describes itself as "an open-source, self-hosted AI coding assistant" with which a team can set up its own code completion server and IDE extensions (Tabby docs). Choose it if you want shared, self-hosted completion in the editor. Continue with Ollama gives "complete control over your AI models, ensuring privacy and enabling offline development workflows" inside the IDE (Continue guide). Aider collects analytics only if you opt in, and aider --analytics-disable opts out forever (Aider analytics). OpenCode discovers Ollama, LM Studio and vLLM on their default local addresses with no account connection (OpenCode providers).

Questions people ask

Is there an AI coding tool that doesn't send our code to OpenAI or Anthropic?

Yes, if you choose the model. CodeAF sends prompts and code only to the provider you connect, so with Ollama, llama.cpp, LM Studio or another OpenAI-compatible endpoint in your network, no code goes to OpenAI or Anthropic.

What does CodeAF send to AgentField?

Anonymous usage counts (version, OS, mode, session counts, errors and token totals as bands) and, when the Model Pool sends, text-free model scores. Neither carries code, prompts, paths or keys. CODEAF_TELEMETRY=off and model_pool = off stop them.

Can CodeAF run in an air-gapped environment?

Not with a documented procedure today. It is one binary, local models need no key and both outbound streams can be turned off, but there is no official offline installer and no published air-gapped test.

Does CodeAF have a self-hosted server for the whole team?

No server product is documented. CodeAF runs on each machine; a team can share one dev box inside the network over ssh with codeaf chat --host.

Where are my API keys stored?

In your CodeAF profile on that machine. A key can also be a variable name like $MIXPANEL_KEY, read fresh each time it is used.

Sources

Last updated 2026-11-21. Checked against CodeAF v0.8.3 docs and README and the Tabby, Continue, Aider and OpenCode documentation on that date.