Now at v0.3 · for VS Code & code-server

The coding agent
you own, end to end.

Baala Code works inside your editor. The agent loop, tools, prompts and provider adapters are yours. Point it at any OpenAI-compatible model and keep every key, file and conversation on your own server.

Talk to us See how it works

uploader.ts — acme-api

uploader.tsretry.ts

12export async function upload(file: Blob) {
13  return withRetry(() => put(file), {
14    attempts: 5,
15    backoff: 'exponential',
16    retryOn: [429, 502, 503],
17  });
18}
19
20// Honour Retry-After when the server sends it
21function delayFor(res: Response, n: number) {
22  const header = res.headers.get('Retry-After');
23  if (header) return Number(header) * 1000;
24  return Math.min(500 * 2 ** n, 30_000);
25}

Baala Code Ready

Add retries with exponential backoff to the uploader, then run the tests.

✓read_filesrc/uploader.ts

✓edit_filesrc/uploader.ts+3 −1

run_commandnpm test

Allow onceThis sessionDeny

Context18%

Your endpoint

Hosted APIs, self-hosted gateways or serverless GPUs: anything that speaks /v1/chat/completions.

Your rules

Edits, commands and reads outside the workspace wait for you. Sensitive files always ask.

Your server

History, profiles and preferences live on the machine running your editor, not in a third-party cloud.

Product

Everything a serious agent needs.
Nothing you can't inspect.

Any OpenAI-compatible model

Switch between endpoints and models from the input bar. Model lists load straight from GET /models.

Permission-first tools

Reads and searches run directly. Edits and commands show a permission card you approve once or for the session.

Checkpoints & Undo

Every turn records the files it touched. Review diffs in the changed-files card and revert them in one click.

Context that lasts

Old tool output is trimmed and long sessions compact themselves near the limit, so work continues without starting over.

Resilient streaming

Stalled streams are detected and dropped connections retried, keeping the partial reply. Rate limits honour Retry-After.

Steer while it works

Keep typing during a turn. Follow-ups are queued and delivered before the next model call. Stop is always one click away.

How it works

A loop you can read,
step by step.

Baala Code runs its own agent loop: model call, tools, results, repeat. Every step appears in the panel as it happens.

read_filelist_dirsearchedit_filewrite_filerun_command

01
Understand
The model reads your request with the project's AGENTS.md and any custom instructions. Reasoning streams into the panel.
02
Act
It reads files, lists directories and searches with ripgrep to build context, then proposes edits or commands.
03
Ask
Anything that changes your workspace pauses on a permission card. Nothing is written or run without your say.
04
Review
Diffs land in the changed-files card. Keep them, or undo the turn and every file goes back as it was.

Safety

Guardrails that
don't switch off.

"Allow everything this session" speeds up routine work. These checks stay on even then.

Sensitive files always ask

Reading or writing .env*, *.key, *.pem, SSH keys or cloud credentials needs explicit confirmation every time.

Workspace boundaries

Reads, listings and searches outside the open workspace need your authorization first.

Deletes always confirm

Shell commands that delete files are flagged and held for approval, whatever the session allows.

Secrets stay out of context

Search skips sensitive files by default, so keys are never indexed into results or sent to the model.

Credentials by reference

Profiles point to a key file or environment variable. The panel never receives an inline secret and cannot write one.

Inspect every request

Open the exact payload of the last model call at any time. The whole system prompt lives in your repository.

Models

Bring any model.
Switch in a click.

A profile bundles an endpoint, its credentials and its models. Keep as many as you like and switch the active one from the panel.

Profiles stored on the server, readable only by its owner
Each profile remembers its last model
Serverless GPUs: waits through cold starts and shows progress, without waking idle endpoints just to check them

Changelog

Shipping steadily.

0.3.2
A quieter panel
The endpoint card folds into a status chip once the model is ready.
0.3.1
Safety guards
Sensitive-file confirmation, workspace boundaries and secret-free search.
0.3.0
Long sessions
Context compaction, resilient stream retry and message queueing.
0.2
Profiles
Switchable endpoints and live endpoint status. Rename and delete conversations.
0.1
First release
Our own agent loop, six tools, permission cards, history and Undo.

Running AI on your own infrastructure?
We'd like to hear from you.

Early access, team deployments and partnerships.

hi@baala.net

The coding agentyou own, end to end.

Your endpoint

Your rules

Your server

Everything a serious agent needs.Nothing you can't inspect.

Any OpenAI-compatible model

Permission-first tools

Checkpoints & Undo

Context that lasts

Resilient streaming

Steer while it works

A loop you can read,step by step.

Understand

Act

Ask

Review

Guardrails thatdon't switch off.

Sensitive files always ask

Workspace boundaries

Deletes always confirm

Secrets stay out of context

Credentials by reference

Inspect every request

Bring any model.Switch in a click.

Shipping steadily.

A quieter panel

Safety guards

Long sessions

Profiles