Agents that work in your repository, under your rules.
Connect Claude Code, Cursor, VS Code, or ChatGPT in one step. Your agents pick up issues, edit code on the server, run your tests and open pull requests. You review and merge.
Connect
A remote MCP server with browser sign-in through OAuth. You choose read-only or read and write, and which repositories, when you approve.
https://sailfork.com/mcp
https://sailfork.com/mcpEvery toolhttps://sailfork.com/mcp/readonlyOnly tools that never change anythinghttps://sailfork.com/mcp/x/{toolsets}Only the toolsets you list, such as pulls,ci
To run the server on your machine as the account you signed in with, use sailfork mcp. Steps for each client are on MCP server.
What an agent does with an issue
From the queue to a pull request you can review.
- 1
Take the next issue
The agent claims it, so nobody else works it. The claim lapses if the agent stalls, after a time you set between 5 and 1,440 minutes, and the issue goes back to the queue.
issue_next - 2
Open a caisson
A caisson is an agent’s own branch and its unsaved edits, kept on Sailfork. The agent reads and writes through the API, so nothing is cloned and no machine starts. Opening one returns the path rules it writes under.
caisson_open - 3
Edit and search
It reads files, searches by text or by syntax, and changes them in place. Edits stay in the caisson until they ship.
caisson_edit - 4
Run the tests
Tests run against the caisson’s uncommitted edits in a throwaway sandbox on your own Drydock machines, then the sandbox is destroyed.
caisson_test - 5
Ship a pull request
One call commits the edits and opens a pull request that closes the issue.
caisson_ship - 6
Get it reviewed
The pull request goes through your branch rules and reviewers like any other. If your policy requires it, a review of the code must exist before it can merge.
caisson_review
Tools it already knows
108 tools in seven toolsets: caissons, issues, pull requests, CI, repositories, sandboxes, and context.
- Reads of a line range, or of up to 100 files in one call
- Edits that match an exact string once, or every match
- Batches of edits that land whole or not at all
- Text, regex and code search
- Search and rewrite by syntax tree with ast-grep, so comments and strings never match
- Language servers that answer references and rename a symbol everywhere
You stay in charge
Rules are set by the owner of the repository, in the browser.
- Allow, ask or never, per path
- Mark a path never and a write to it is refused. Mark it ask and the write waits for a person.
- Approvers are told
- People who can approve a held write get an inbox notice and an email.
- A token cannot decide
- No token can approve a held write or change the policy. Only a person in the browser can, so an agent cannot answer its own request.
- Every commit names its source
- Agent pull requests and commits are marked as agent work and name the person whose token was used.
- Undo
- Any agent action can be undone once within 24 hours with sailfork undo.
- A read-only server
- Connect an agent to the read-only address and it can look at everything and change nothing.
What it costs
A sandbox bills for every second it runs, even while the agent waits on a model. A caisson is a branch and a table of edits on our servers, so nothing runs between calls.
- Open a caisson
- 100 µs
- Read a file
- 1 µs
- Write a file
- 10 µs
- Commit
- 500 µs
100 agents working 8 hours a day: $132 a day in sandboxes on E2B or Daytona, $0 in compute on caissons.
An upper bound that assumes the sandboxes bill the whole time. Compute starts on a caisson only when an agent runs tests or a language server.
| Service | One hour, 2 vCPU and 4 GiB | While idle | To start |
|---|---|---|---|
| Caisson | $0 compute | Nothing runs | 10,000 agent changes a month |
| Sailfork sandboxes | On your machines today | Your machines | Hosted sandboxes coming |
| Northflank | $0.07 | Not published | $20 a month of resources |
| Vercel Sandbox | $0.08 to $0.34 | Memory still billed | 5 CPU-hours a month |
| Cloudflare | $0.08 to $0.22, 8 GB | Nothing once asleep | $5 plan, 375 vCPU-minutes |
| Fly Sprites | Up to $0.16 | Storage only | $1 credit |
| E2B | $0.17 | Compute stops on pause | $100 credit |
| Daytona | $0.17 plus disk | Disk only | $200 credit |
| Blaxel | $0.17, 4 GB | Snapshot storage | Up to $200 credit |
| GitHub Codespaces | $0.18, 8 GB | Storage only | None for organizations |
| Modal | $0.24 | No pause | $30 credit a month |
| Runloop | $0.32 | Storage only | $50 credit |
Vendor rates as published, checked November 12, 2026, at the closest size each sells. Ranges depend on CPU use. The day figure is the hourly rate multiplied by the agents and hours, not a measurement. Timings are medians from a Mac in San Francisco: open against sailfork.com, the rest on a release build with object storage.
Agent changes
An agent change is one write an agent makes through a caisson. Every plan includes 10,000 a month, then $2.00 per 1,000. Pushes over git never count.
Storage for your repositories
10 GB
Git history as stored after compression. A checkout on disk is usually larger.
Pro
50 GB included
- Pro plan
- $12
A month, once billing starts $12
Today, during early access: $0.
Tested with real agents
11 tasks, 3 runs each, Claude Haiku through Claude Code: 33 of 33 passed over MCP, and 33 of 33 over the CLI. How we test agents
Made for agents to read
Every command prints JSON on request, and output to an agent is capped at 64 KiB.
- llms.txt
A map of the docs for a model
- MCP server card
What the server offers and how it signs in
- auth.md
How an agent signs in
sailfork commands --jsonEvery command, flag and whether it changes anything
sailfork schemaThe JSON shape of a command’s output