Step 1
Enroll a host
Point Vise at a machine you control. It claims sessions and runs the agent.
$ vise hosts create laptop
Vise runs coding agents on hosts you control.
Open-source background coding agents. Describe a task and a repo. Vise runs the agent on your machine, on your model credentials, and tracks the pull request until a person merges it.
curl -fsSL https://vise.sh/install | sh
macOS (Apple Silicon) · Linux (x86_64, arm64) · Docker and git are the only prerequisites
A person still reviews and merges every PR. Vise just doesn't stop watching after it opens.
A session, end to end
Watch one run.
Every state change is a session event:
pr_state_changed, checks_state_changed.The problem
The agent isn't the hard part anymore.
Running one well is. It takes a host, a way to launch it from somewhere other than your laptop, and someone watching the pull request after it opens. Most teams that run agents seriously end up building all three by hand. Some of the most engineering-mature teams already have. You'd rather not maintain your own fork of it.
The comparison
Vise vs. hosted background agents
Hosted background agents run the vendor's agent, in the vendor's cloud, on the vendor's terms. Vise is the same job on your infrastructure.
| Vise | Hosted background agents | |
|---|---|---|
| Where the agent runs | Your host: laptop, VM, CI runner | Vendor's cloud |
| Whose model credentials | Yours. Vise is never in the inference path | Vendor's, billed through them |
| Which agent harness | Claude Code today; more can be added | Vendor's own agent |
| Repo access | Clone on your host, your GitHub token | Vendor app installed on your org |
| Who merges | A person, always | Varies; some auto-merge |
| PR tracking after open | Polled to merge; review and check state recorded as events | Varies |
| Can you read the scheduler and sandbox? | Yes, it's open source | No |
| Hosted option | No. Self-hosted only | Yes |
How it works
Three steps. No changes to the agent you already run.
Step 2
Start a session
Describe the task and the repo. The host clones it and runs the agent there over ACP, the Agent Client Protocol, an open standard for driving coding agents.
$ vise sessions create "fix the flaky retry test, run the suite, open a PR" --repo acme/billing --watch
Step 3
Vise tracks the PR
The agent branches, commits, pushes, and opens the PR itself. Vise polls it and records review and check state until it merges. When a reviewer asks for changes, a follow-up session pushes the fix to the same PR.
$ vise sessions follow-up ses_8f31 --watch
Configure the agent
One API call sets the whole session.
# POST /sessions
{
"agent": {
"harness": "claude-code",
"model": "<your model>",
"instructions": "follow CONTRIBUTING.md, keep PRs small"
},
"environment": {
"kind": "github_repo",
"repo": "acme/billing",
"base_branch": "main"
},
"input": "fix the flaky retry test, run the suite, open a PR"
}
HarnessThe agent that runs the session. Set per session.
ModelSet per session. The agent runs on the model and credentials you configure.
InstructionsStanding guidance for the agent: conventions to follow, what to avoid.
EnvironmentThe repo and the base branch the work starts from.
What you get
The infrastructure, not another agent.
Self-hosted
vise-server needs Postgres. vise-host runs on any machine you enroll: a laptop, a VM, a CI runner.Built on ACPVise drives the agent over ACP, not a hardcoded integration. Claude Code is the harness wired today. More can be added without changing the protocol.
PR trackingEvery PR Vise opens is polled until it merges or closes, with review state and check state recorded as events.
Follow-up sessionsWhen a reviewer asks for changes, a follow-up session gets the review thread and pushes a fix to the same PR.
Your model credentialsThe agent runs on credentials you configure. Vise never sits in that path.
Open sourceRead the scheduler, the sandboxing, and the PR tracker before you trust them with your repo.
Fit
Built for teams already running agents for real
- Agents already run against a real repository, and more than one person depends on the result.
- Someone on the team has already written a script or a cron job to launch agents outside a laptop.
- You'd rather read the code that runs your agent than take a SaaS's word for it.
Not a fit if
- You want autocomplete or an in-editor pair programmer. Vise runs the agent you already use; it doesn't replace your editor.
- You want a hosted product with nothing to run yourself. Vise is self-hosted: you run Postgres and the host.
- You need code generation without human review. Every Vise PR waits for a person.
Why open source
Read it before you trust it.
The thing that runs an agent against your repo should be something you can read. The scheduler, the sandbox, and the PR tracker are all in the repository.
Vise started inside our own tools. They needed coding agents running somewhere, and we didn't want to tie them to one model provider, one harness, or one environment. So we built the neutral layer underneath and open-sourced it.
Status
What ships today, what's next
| Today | Next |
|---|---|
| Claude Code harness · GitHub repositories · PR and check tracking · follow-up sessions · macOS and Linux hosts | mcp_servers wired to the agent (the API accepts the field today) · additional harnesses |
Public roadmapsoon · Contributing
FAQ
The questions engineers ask
Do I have to change the agent I already use?
No. Vise drives the agent over ACP. Claude Code works today, and you set the harness per session.
Where does the agent run, and what can it reach?
On a host you enroll. The agent runs in a clone of the repo on that machine, using the GitHub token and model credentials you configured.
Who opens the pull request?
The agent does, from a branch pushed from your host. It authenticates with the GitHub token you configured at install, so commits and the PR are authored by that token's user. A person reviews and merges. Vise does neither.
Is any of this hosted by Vise?
No. You run vise-server, Postgres, and vise-host yourself. Nothing in the loop is a Vise-run service.
What's the license?
Dual-licensed under MIT or Apache-2.0, your choice. The server, the host, and the CLI are in one repository.
What's ACP?
The Agent Client Protocol, an open standard for driving coding agents. Vise speaks ACP instead of a hardcoded integration, so a new harness is a change to the host, not a fork of Vise or a new API.
Run it on a machine you already have.
curl -fsSL https://vise.sh/install | sh
macOS (Apple Silicon) · Linux (x86_64, arm64) · Docker and git are the only prerequisites