Last updated · Santosh Kumar Radha

An Apache 2.0 coding agent with approvals for regulated teams

CodeAF is Apache 2.0 licensed. How its approvals, budgets, isolated branches and records work for enterprise and regulated teams, and where the gaps are.

CodeAF is an open-source coding agent by AgentField under the Apache 2.0 licence, which grants commercial use and an express patent licence from contributors; for a regulated team it adds approval modes, actions that always ask, work isolated on its own branch, spend limits at every level and per-task records you can inspect. It does not have a certification, a central admin console or an organisation-wide audit log export, and this guide says where those gaps are and what covers them.

At a glance

ControlWhat CodeAF doesSource
LicenceApache 2.0repo, licence text
Approval modesAlt+A walks asks, guardian, YOLO; new conversations start in YOLO unless you saved another choicecommands, FAQ
Always asksA short list of dangerous commands, and anything sent in your name (email, Slack, calendar events)FAQ, connections
Unattended workRuns only under approvals already in your profile; anything that would ask refuses and waitsrunning unattended
Where changes landOwn copy of the folder; kept on its own branch when you are on mainFAQ
SpendEvery call priced; daily, conversation, task and team limitsbudgets
Recordscodeaf why <task-id>, codeaf logs, saved specialist reportscommands
CertificationNone documented (no SOC 2, no HIPAA attestation)
Central admin policyNot documented

The licence

Apache 2.0 lets your company use, modify and redistribute CodeAF commercially, and its section 3 grants a patent licence from each contributor for their contributions (Apache License 2.0). CodeAF itself is free; you pay your model provider (FAQ). Have your counsel read the licence; this page is not legal advice.

Step 1: set the approval mode before anyone uses it

A new conversation starts in YOLO unless you saved another choice: ordinary commands run without asking. Press Alt+A to walk to asks (or guardian) in each conversation that needs it (FAQ, commands). Whatever the mode, a short list of dangerous commands and anything sent in your name always asks first. Each connected account has two switches in /settings: read, which is on, and act in your name, which asks first (connections).

Managers need the same care. On the default YOLO a manager starts new members at once; on asks each start waits on a card such as ◆ manager wants to start @platform. Set this in each manager's conversation. Permission prompts never go to a manager: they come to you (managers).

Step 2: keep changes off protected branches

Handed-off work runs in its own copy of your folder, so your checkout does not move. At the end it is merged into your folder, or kept on its own branch when you are on a branch such as main that tasks never merge into (FAQ). senior-dev cuts a branch named after the task, such as task/rename-loadconfig-to-loadsetting-7d6220, and runs your build and tests on the frozen tree; read the verification result before merging (senior-dev). Your normal pull-request review and branch protection stay the gate. For a second reader, /review owner/repo#123 reports findings and posts nothing until you ask; a partial review cannot be posted as an approval (review, PR review guide).

Step 3: hold unattended work to what you already allowed

Schedules and watches run with nobody at the keyboard, so nothing widens what they may do: a run works under the approval rules in your profile, "not under any file checked into the repo", and anything that would have asked you refuses instead (running unattended). Each standing order carries backstops of $5 a run and 10 runs a day unless you named your own. Headless jobs in CI follow the same idea: codeaf do takes no --yolo and refuses a plan that would cross your daily limit unless you pass --yes-spend (headless, CI guide).

Step 4: know what the records are

  • codeaf why <task-id> says what a piece of work did and what it cost; codeaf logs --tail <n> shows recent model calls (commands).
  • Every task names its crew and cost, and spend pages break cost down by model and role.
  • /sec saves Markdown, JSON and SARIF reports; /review saves review-report.md and review-report.json, including the commit reviewed (specialists).
  • /standing lists every standing order with its trust count, such as earning trust 3/5. The count is a label, never a permission.

These are records on the machine that ran the work. They are not signed, centrally collected or exported to a SIEM by any documented CodeAF feature. For organisation-wide agent identity, policy and a signed audit trail, AgentField (the control plane from the same team) is the layer built for that: see agent identity and audit trails.

When something else fits better

GitHub Copilot logs cloud-agent activity on GitHub to the enterprise audit log (search actor:Copilot), retains events for 180 days, and recommends streaming to a SIEM; GitHub notes the log "does not include client session data, such as the prompts a user sends to Copilot locally" (GitHub docs). Choose it if your audit evidence already lives in GitHub Enterprise. Claude Code lets administrators deploy managed settings that user and project settings cannot override, including disableBypassPermissionsMode (Claude Code permissions). Choose it if you need centrally enforced permission policy today. Codex can export OpenTelemetry logs of prompts and tool approvals and results, off by default and enabled through an [otel] table in config.toml (Codex config).

Questions people ask

Is there an Apache licensed AI coding agent for enterprise use?

Yes. CodeAF is Apache 2.0, a terminal coding agent you run on your own machines with the models you choose. It is labelled an early preview in its README, so plan a pilot rather than a fleet rollout.

Does CodeAF ask before it runs commands?

By default ordinary commands run without asking, because new conversations start in YOLO. Alt+A switches to asks or guardian, and dangerous commands and anything sent in your name always ask in every mode.

Is CodeAF SOC 2 or HIPAA certified?

No. No certification is documented for CodeAF. Its controls (approvals, branch isolation, budgets, local records) can support your own evidence, but they are not an attestation.

Does CodeAF have an audit log?

It keeps per-task records on the machine: codeaf why <task-id>, codeaf logs and saved specialist reports. There is no documented central or signed audit log export; AgentField covers signed audit trails across agents.

Can an admin enforce approval settings for every developer?

Not through a documented CodeAF feature. Approval choices are saved per profile and per conversation. If central enforcement is a hard requirement, plan for it outside CodeAF.

Sources

Last updated 2026-11-21. Checked against CodeAF v0.8.3 docs and the GitHub, Claude Code and Codex documentation on that date.