mrkeyoor.com_
Fri 06 Nov 03:55 UTC
Automationevaluationupdated 06 Nov 2026

switch review

Agent Switch connects people and their existing AI agents in shared rooms linked to Slack, Teams, Discord, Telegram or Mattermost. It lets a team bring agents into conversations without replacing its chat app or choosing a new model provider. The source is available under Apache 2.0 with a modified Commons Clause restriction.

Verdict

Agent Switch built in 4 seconds in our sandbox, but tests stopped before execution because pytest_asyncio was missing, leaving shared-room operation unverified. Try it when your team already has agents and wants them available in existing chat apps. Require a remote-host recovery check and a room-permission trial before depending on it for unattended work.

We ran it

Lab card: what happened when we ran switchScreenshot of switch (www.switchagents.ai)
Install✓ · 15s118 packages · 170 MB
Build✓ · 4s
Tests✗ · 2sran, no count parsed
Known vulns0(pip-audit)
Repo4768 files~899,020 lines of source · 51.4 MB · 9 CI workflows · tests dir

Answers from our run

Does switch build from source?

Dependencies installed in 15 seconds (118 packages), and the build succeeded in 4 seconds. We cloned commit da5336c into a clean Debian container with 3 CPUs and no project-specific setup.

Do switch's tests pass?

No. The project's own test command exited unsuccessfully after 2 seconds, and its output did not expose an individual test count.

Does switch have known vulnerabilities in its dependencies?

pip-audit found none in the dependency tree at the time of our run.

Who should not use switch?

Teams expecting Switch to supply an agent or model: the README explicitly says it ships neither.

What are the alternatives to switch?

Open WebUI, n8n. Agent Switch built in 4 seconds in our sandbox, but tests stopped before execution because pytest_asyncio was missing, leaving shared-room operation unverified.

SetupN/AThe core install and build completed, but desktop onboarding, provider sign-in and a supported remote deployment were not assessed.
DocsN/AThe README and remote-host guide give specific boundaries, but their instructions were not reproduced across supported platforms.
CommunityN/ARecent push and pull-request activity was checked, but response times, adoption and maintainer support capacity were not assessed.
MaturityN/AThe core test command stopped on a missing import, and no shared-room workload or recovery exercise established operational reliability.

Who it’s for

Engineering teams that already run coding agents and want colleagues to address them in shared chat rooms.
Operators willing to manage a Switch server, agent-provider sign-ins and the permissions of each remote working directory.
Teams evaluating shared agent context across existing messaging platforms, with a limited pilot before broader access.

Who it’s NOT for

Teams expecting Switch to supply an agent or model: the README explicitly says it ships neither.
Operators requiring unattended recovery after a remote host reboot: the documented Console-managed server needs a manual start.
Teams whose remote hosts must run Windows: the remote-host guide says there is no Windows installation path. This does not describe the Windows desktop client.
Anyone needing an unrestricted Apache-only license: the license adds a modified Commons Clause restriction on selling the software or derivatives.

Setup reality

Our November 5, 2026 sandbox run at da5336c installed the core project in 15 seconds with 118 dependencies and a 170 MB footprint, then built it in 4 seconds. Tests exited 4 after 2 seconds: loading core/tests/conftest.py failed because pytest_asyncio was missing. pip-audit reported 0 known vulnerabilities, which is not a security guarantee.

Switch requires an existing agent and model arrangement. Remote hosting needs working SSH access, a supported host and Docker for the server. You sign in to the agent provider on that host yourself and supply its working directory. The desktop onboarding route was not exercised by the core run.

The configured sandbox had 3 CPUs, 8 GB RAM and image ghcr.io/astral-sh/uv:python3.12-bookworm. That label does not establish the actual runtime or OS version. Remote-host docs exclude Windows and say a host reboot requires manually restarting the server. Console telemetry is on by default, while Core telemetry defaults to off.

Shared rooms need agents you already operate

Agent Switch brings existing AI agents into Slack, Teams, Discord, Telegram and Mattermost. Its README says it ships no agents and no models. If your team already uses a coding agent, the attraction is letting a colleague join the conversation and see the context instead of asking someone to paste a private session into chat.

That makes Switch relevant to a team with an agent workflow already worth sharing. A developer who only wants a private model chat has little reason to add rooms, messaging integrations and remote agent hosts. The system connects parts you still have to choose and operate.

The README describes room membership as the boundary for an agent's context. It also describes roles, instructions and tools attached to agents. Those are product claims we would want to check with restricted accounts. Our recorded core run did not establish that room isolation or any messaging connector worked.

What happened when we ran it

Our Agent Switch core installed in 15 seconds and built in 4 seconds at commit da5336c on November 5, 2026. The run recorded 118 dependencies and a 170 MB installed footprint. Tests exited 4 after 2 seconds.

The supplied tail identifies the immediate failure: loading core/tests/conftest.py reached an import of pytest_asyncio and raised ModuleNotFoundError. The test command stopped during that loading step. It provides no passing-test result, and it cannot tell us whether the core's behavior would pass after the missing module becomes available. The record does not establish why that module was absent.

The configured sandbox had 3 CPUs, 8 GB RAM and image ghcr.io/astral-sh/uv:python3.12-bookworm. The image name cannot establish the actual Python or operating-system version without toolchain logs. The measured Python project was in core/, so its build result does not validate desktop installers or a connected agent session.

pip-audit reported 0 known vulnerabilities. That finding concerns known advisories in the recorded audit, not authorization design or agent access to your files. The scan also recorded 9 CI workflow files and a tests directory, with no Dockerfile detected. That scan does not prove published server containers are absent, and CI file presence does not prove a passing pipeline.

A closed laptop and a reboot are different cases

Switch's remote-host guide says an agent on another machine can keep answering after you close your laptop. It also says a host reboot stops the server and you must start it again yourself. That is a material limit for a team expecting an agent to answer while its operator sleeps.

The documented host path needs working SSH access and an account allowed to install software. A server host also needs Docker. The guide lists Linux or macOS and explicitly excludes a Windows remote-host installation path. The README's Windows desktop download serves a different role, so its existence does not remove that host restriction.

Console installs the provider's command-line tool, but the operator must sign in on the host. It also leaves cloning the repository to you. The agent uses the code in its remote working directory, which can differ from what is on your laptop. Before asking a colleague to rely on an answer, check that directory's revision and the access available to the host account.

For a pilot, close Console and confirm the agent remains reachable through the intended chat app. Then check the documented reboot recovery separately. Those are proposed acceptance checks, not workloads completed in our sandbox.

Desktop telemetry starts enabled

The README says Switch Console collects anonymous usage analytics by default, with a first-run notice and a Share usage data toggle. Listed fields include event names, application version, operating system, provider category and an anonymous installation ID. The project says it excludes prompts, source code, credentials and model outputs. We did not inspect network traffic.

Core has a different documented default: server usage telemetry is off unless enabled. A Console-managed server follows additional rules described in the README, including applying a server setting at startup. Check both components when evaluating your organization's reporting policy. Self-hosting alone does not establish that the desktop sends no events.

The license includes a sale restriction

The license combines Apache 2.0 with a modified Commons Clause condition. Its text excludes the right to sell the software or derivatives and defines sale to include certain paid hosting or related services. Treat it as a license with an added restriction when comparing deployment options. Calling it Apache-only would omit a consequential term.

For a different workflow, Open WebUI documents connections to Ollama and OpenAI-compatible APIs in its own interface. It is a candidate when direct model access is the requirement. n8n combines a visual workflow canvas with custom code and describes itself as fair-code. Consider it when the automation should be designed as an explicit workflow. Neither alternative was benchmarked against Switch.

Active development leaves operational questions open

The repository's last push was November 6, 2026, with recent pull requests concerning agent presence and Console login-error handling. That indicates ongoing work, not verified fixes in a released desktop package. We leave numerical ratings unassessed because the core checks and public documents do not establish deployed reliability, support capacity or platform-wide setup quality.

Start with one restricted room and an agent working on a disposable checkout. Keep access narrow until a colleague can use it, the host recovery procedure is understood, and the missing test prerequisite has been investigated.

Alternatives

ProjectWhat it isPick it when
Open WebUI gh↗A self-hosted AI interface that connects to local models through Ollama and to OpenAI-compatible APIs.Pick this instead when you want a dedicated place to use models rather than bringing existing coding agents into team chat rooms.
n8n gh↗A workflow and AI-agent platform combining a visual canvas with custom code.Pick this instead when explicit workflow design is the main job. Its README identifies it as fair-code and documents separate license terms.

What people are saying

  1. [theverge] Nintendo says it’s not planning another Switch 2 price hike
  2. [github-trending] sandbox-quantum/switch
  3. [theverge] Nintendo is losing a Switch cartridge supplier
  4. [hackernews] Landline phones will switch off across Britland in 100 days, no extensions
  5. [hackernews] Switch Lowering in GCC
  6. [huggingface] Granite Switch 4.1.1 — Update models like you update software

Sources

  1. Agent Switch README
  2. Remote-host operation
  3. Agent Switch license
  4. Open WebUI README
  5. n8n README
  6. Switch activity

More automation reviews

Horizon · ai-goofish-monitor · agent-desktop · Microsoft-Rewards-Script · moli · MatrixMedia · the whole board →