LoopForge: picking a repo does not narrow the token.
The coding-agent workspace separates a repository choice from the GitHub grant behind it. Check both before connecting valuable code.

Concept photo: programming code on a screen, by luis gomes / Pexels. Illustrates programmable tools, not a LoopForge product screenshot. Used under the Pexels license, which permits article, RSS and social reuse. No endorsement implied.
A developer picks one repository, asks for a bug fix and waits for a review. That looks like a small assignment. The GitHub permission behind the connection may cover much more than the repository on screen.
That distinction matters in Tencent's LoopForge, a self-hosted workspace for coding agents receiving fresh attention in November 15 coverage. Its current deployment guide makes an unusually useful disclosure: selecting a repository binds it to a project, but does not narrow the GitHub OAuth grant.
The picker and the permission are different controls
LoopForge's documented connection uses an OAuth App requesting the repo scope for private cloning, pushes, issues and pull requests. Its repository picker can show repositories owned by the account, shared with it or available through organizations. Choosing one does not turn that broader authorization into a one-repository credential. Organization restrictions may still require separate approval.
GitHub's authorization documentation explains that device flow obtains an access token after the user approves. It also distinguishes GitHub Apps, which offer finer permissions and more repository control, from OAuth Apps. An easy sign-in flow is not evidence of a narrow grant.
LoopForge adds another distinction worth checking before a team connects accounts: its deployment guide says issue import and pull-request status synchronization use the project owner's Git token, while Git operations and the optional GitHub MCP integration use the responsible user's token. The person clicking and the account supplying access are not always the same.
A review gate cannot shrink a credential
The project README describes workflows built from agent, script and human-gate steps, with artifacts and logs for review. Those can make a software delivery easier to inspect. They do not, by themselves, reduce the permissions granted earlier on GitHub.
The same documentation targets trusted teams, not separate untrusted tenants. The deployment guide says every signed-in account can list projects, requirements and runs. It also warns that the default container setup gives the platform access to the host Docker daemon. Neither a project label nor the word sandbox should substitute for reading those boundaries.
Start with the no-key check
For a first evaluation, use the README's script-and-approval demo before connecting a valuable repository or model account. The documented demo needs neither a model key nor a repository. Check which artifacts the reviewer sees, what approval continues and which account would supply access during real work.
Then inspect the actual GitHub authorization screen and organization policy. If the proposed access is broader than the job permits, stop before granting it. A repository choice inside an application is not a replacement for a suitably limited account or credential.
This is a documentation review, not a GigSoul deployment or security test. LoopForge's published verification record describes project-run checks and explicitly limits what they prove about an operator's external deployment. The useful takeaway is smaller than a safety verdict: approve the work and examine the access separately.


