AI agent
A bounded, tool-using agent that plans, edits files, runs builds, reads logs, and repairs errors.
Every step is visible: the plan, each tool call, the files it changed, the token usage, and the snapshot it took before starting.
The agent is deliberately constrained. It can only touch files in your project, only run allow-listed commands (install, build, test, type-check) in the sandbox, and only retry a failing validation twice before stopping and explaining what it found.
What you get
Three modes
Build creates and edits; Ask explains without changing files; Fix targets a failing build or a specific error.
File tools
Read, write, rename, and delete project files. Each call is streamed to the chat as it happens.
Sandbox commands
Allow-listed install, build, test, and type-check commands run in the isolated runtime, never in the web process.
Bounded repair loop
On a failed validation the agent reads the error output and attempts a repair at most twice.
Snapshot and revert
A snapshot precedes every run. Review the run's diff and revert it with one click.
Provider abstraction
Anthropic and OpenAI models through one interface; the operator chooses the default provider and model.
How it works
Step 1: Plan
Reads the prompt and relevant files, emits a numbered plan.
Step 2: Edit
Applies file changes with tools; each change is listed in the run.
Step 3: Validate
Runs type-check, build, or tests in the sandbox.
Step 4: Repair
Reads errors and fixes them, bounded to two attempts.
Step 5: Report
Summarises what changed and what it could not do.
Limitations and requirements
- AI features require an Anthropic or OpenAI API key configured by the operator of your Codexa instance. When no key is configured the agent shows an explicit "not configured" state instead of pretending to work.
- The agent cannot browse the web, access your other projects, or run arbitrary shell commands.
- Token usage is recorded per run and counts toward plan limits; long files cost more. See credits.
- Runs can be cancelled, but a cancelled run may leave partial edits. Use the run's revert action to restore the snapshot.
Frequently asked questions
Which models are used?
Whatever the operator configures: an Anthropic model or an OpenAI model. The provider and model are shown in the agent's status panel.
Is my code sent to the model provider?
Yes, the files relevant to a request are sent to the configured provider to generate the response. See the AI policy for details and how to avoid sending secrets.
Start building with Codexa
Free plan, no card required. Describe your idea and get a real project you can edit, run, and deploy.