Should you trial Rakazo?
Trial Rakazo when retaining a bot's working files and browser state is worth operating a database and computer service. Postpone adoption if you need verified deployment or recovery results before connecting important accounts. Our run on 2026-10-17 at commit e55263f installed and built the source, but did not verify a working assistant deployment.
The README at our measured commit describes persistent bots with conversations, memory, routines and history, plus browser, terminal, file and graphical desktop access. Web, Electron and mobile applications connect to the same API. These are documented capabilities, not completed workloads from our sandbox.
What our sandbox established
Installation succeeded in 42 seconds and build succeeded in 21 seconds; the test command exited 1 after 332 seconds. Our harness stores each command's outcome, exit code, elapsed time and timeout status separately from its parsed test counts. None of these steps timed out.
The attached stored parser summaries record 8,605 passed, 26 failed and 256 skipped out of 8,887. Treat those figures as recorded parser summaries, not a verified count of exhaustive unique tests. No new execution or complete-log reconciliation was performed for this correction.
The raw tail identifies failures in packages/testkit/src/playwright-artifact-publication.test.ts and .agents/skills/pr-watch/pr-digest.test.ts. Those suite labels locate the reported failures; they do not reveal their causes. The supplied tail omits the underlying exceptions, so it cannot establish whether the failures came from product behavior, fixtures or sandbox conditions.
The tail also marks PostgreSQL-related cases, computer-use acceptance, Docker replay and Team desktop cases as skipped. Skipped cases are not failures, and they provide no evidence that those integrations worked. Our run did not establish live model calls, desktop control or recovery.
The sandbox recorded 1,862 dependencies and 1,841 MB on disk. Its installation log identifies Node.js 22.23.2 on Linux x64; the image label alone would not establish that runtime. No audit result is attached, and scan coverage is unknown.
The MrKeyoor testing methodology explains how command outcomes and parsed counts reach a review. That page provides provenance context; its current procedure does not prove which methodology version governed this October run or expand what the stored record tested.
Choose the installation route deliberately
The pinned README documents two routes. Published images require Docker Engine, the Compose plugin, curl and OpenSSL, without a source checkout or Node installation. Source development requires a supported Node.js version, pnpm 9 and Docker, followed by database configuration and independent authentication, encryption, screen-proxy and sandbox-supervisor secrets.
Our measurements cover source installation, build and tests. They provide no installation timing or deployment result for the published images. A successful source build also does not confirm that the computer image was built or that onboarding reached a working bot.
Use the instructions for the revision you intend to deploy. The pinned README configures a PostgreSQL password and matching database URL; the current README documents separate application and migration credentials. Mixing those instructions would obscure which setup you are evaluating.
Separate computers provide the isolation boundary
Rakazo Team bot folders do not isolate workloads. The current runtime documentation says every Team bot can access the full Team workspace. Separate starting directories and persistent Chrome profiles organize work and browser state, while the bots share the computer's OS user and access.
For workloads requiring isolation, the documentation directs operators to use separate computers. Screen leases coordinate tool calls; they do not isolate mutually untrusted processes. This is a documented operating constraint, not a security boundary we tested at e55263f.
Checkpoints do not replace historical backups
Rakazo's documented local workspace store keeps the latest copy rather than an immutable revision archive. Replacement machines import the latest stored workspace, including portable browser state. That can support continuity after replacement, but it does not provide an earlier version after an unwanted edit or deletion.
The runtime documentation requires a persistent volume, encryption at rest and off-host backups for deployments using that local store. It also distinguishes the portable workspace from a disk snapshot: system packages installed outside the workspace do not move to another provider.
Before relying on persistence, create a disposable file and browser session, replace the computer through your intended lifecycle, and inspect what returns. Separately test restoring an earlier file from your independent backup. These are recommended adoption checks; our sandbox did not complete them.
Alternatives depend on how you want to work
Hermes Agent documents a terminal interface and messaging access through a single gateway process. Its README makes it a relevant alternative when terminal work and messaging are your main entry points. That establishes a reason to shortlist it, not a performance or reliability advantage over Rakazo.
OpenClaw connects its Control UI, CLI, TUI and messaging channels through a Gateway. Its README also describes companion apps and nodes for device capabilities. Consider it when access through existing chat services drives your choice.
These are alternative assistants to evaluate for those workflows, not required Rakazo components. The supplied extracts do not establish equivalent workspace recovery behavior, and we have no comparative deployment results.
How to read the assessment
Setup ease, documentation, community and maturity remain unassessed. Source command results do not establish complete setup ease; documentation extracts do not establish that their instructions produce a working deployment. Repository activity alone cannot establish support quality, and this run does not establish sustained operation or recovery. An unassessed dimension carries no numerical rating.

