Decide which project you are evaluating
The current README introduces MonkeyAI as a team AI workspace for office work, programming and design. The repository is still chaitin/MonkeyCode, and README.monkeycode.md retains an English introduction to the original MonkeyCode development platform. These documents describe different product contexts.
Our run measured ./backend/. The current workspace's operations guide is under monkeyai/. Record the source directory, revision and deployment instructions together before starting a trial. The measured backend result does not establish that the current MonkeyAI workspace works with your models, clients or member permissions.
The current README and operations guide are in Chinese. The retained English introduction helps explain the original coding platform, but it does not replace the current workspace's deployment instructions.
What happened when we ran it
At commit ad169c9 on November 7, 2026, installation succeeded in 50 seconds with 402 dependencies recorded. The backend build succeeded in 197 seconds.
Tests exited 1 after 98 seconds. The supplied Go summary reports 57 passed and 2 failed of 59. The visible tail lists successful package entries, including modelusage, netguard and taskflow, then ends in FAIL. It does not identify the failing cases or show their errors. We cannot attribute the failure to credentials, networking, a database or a product defect from that tail.
The sandbox configuration recorded 3 CPUs, 8 GB RAM and image label golang:1.24-bookworm. Actual runtime and operating-system versions require the raw toolchain logs. The scan recorded 3,897 files, approximately 815,585 source lines and a 91.6 MB checkout. It detected 6 CI workflow files and a Compose file, but no Dockerfile or tests directory. Those detector results are not a complete inventory of deployment components.
Investigate the failed test command before adopting the measured backend. This run did not demonstrate a browser client, model connection or completed workspace task. It measured no office-document quality, coding success rate or interactive performance.
Check permissions and service destinations
The current README documents shared models, knowledge bases, skills, experts, MCP connectors and rules. It also describes resource access controls, usage auditing and accounting. These are documented capabilities; our run did not verify them in an authenticated workspace.
For an acceptance trial, give an administrator and an ordinary member different resource access. Ask the member to complete a small task with an approved model and tool, then inspect the result and recorded usage. Also check that the member cannot use a resource outside that account's permissions. These are proposed checks, not tests completed for this review.
The README says content used with external models or tools still goes to the configured services. Identify those destinations before uploading internal documents. Private deployment alone does not establish that every request remains on your server.
Plan storage and recovery before deployment
The operations guide describes waiting for PostgreSQL, applying migrations and starting Backend and Admin. Operators must change the database and initial administrator passwords. The public address serves the administration page and API and is used for OAuth callbacks, metadata and login redirects. The guide calls for HTTPS in production.
PostgreSQL is only part of the storage plan. The guide says skill ZIP files and connection icons are stored in a private RustFS bucket. It requires backups to preserve both a PostgreSQL snapshot and the RustFS objects referenced by that snapshot. Runtime identity and signing material persist in a separate runtime volume, which the guide also says to back up. We did not deploy or restore these components.
The free-edition table limits teams to 10 seats and excludes technical support and cloud computers. The current client table marks Android, iOS and Linux ARM64 as forthcoming. Choose a client route from that table rather than carrying over the original MonkeyCode introduction's mobile claims.
Keep the measured revision separate from the release
The supplied latest-release record identifies v26072801, published July 28, 2026. The release page identifies its commit as fd025f9. Our measured revision was ad169c9. The supplied repository metadata records a last push on November 6, 2026, but activity after a release does not establish which current workspace instructions apply to that older tag.
Do not treat this review as a deployment test of v26072801 or as proof that current main-branch documentation applies to it. Select and record the revision you intend to operate, then check its own deployment files.
Alternatives to evaluate
Open WebUI documents connections to local models through Ollama and to local or cloud models through OpenAI-compatible APIs. Consider it when model access is the main requirement and an English README would help the initial evaluation.
LibreChat labels attached workspaces highly experimental. Its README also documents sharing agents with specific users and groups. Consider it when shared chat and assistants match the job, and assess attached code workspaces separately. Neither alternative was benchmarked against MonkeyAI here.
Before inviting colleagues, record the source path, revision, deployment instructions and client you intend to support. Then verify a model request and a permitted tool action through an ordinary member account, check denied access and inspect usage records. Those checks address the workspace decision that the backend build leaves unanswered.