Loa/Workspace
English
  • English
  • 简体中文
  • 日本語
  • Español

Loa Workspace

Give an agent a repo, not your machine.

Limit the reach of AI with isolated environments. Workspace creates sandboxes for agents on hardware you own, or on a VM. An agent gets a set of tools over MCP, scoped to one repository, so it can read, search, edit, run commands, commit and push. Commands stay inside the sandbox, and secrets stay out of the model's reach.

Available on Mac and Linux, and as an experiment in Amazon Bedrock AgentCore.

Coming soonSee an example
agent → loa-workspace
  1. The taskThe project is in the workspace, reached by the loa-workspace tools, not in the local directory. Run the quick tests (./test.sh quick), find why any test fails, fix the code, run the quick tests again, and commit the fix with a message that says what was wrong.
  2. 01workspace
  3. contract: workspace/1
    root: /private/tmp/wtx/proj
    receipts: true
    checks: true
    …
  4. 02shell command: ./test.sh quick
  5. exit: 1
    FAIL  /private/tmp/wtx/proj/calc.sh add 2 3: got 'sum = -1', want 'sum: 5'
    FAIL  /private/tmp/wtx/proj/calc.sh add 0 0: got 'sum = 0', want 'sum: 0'
    FAIL  /private/tmp/wtx/proj/calc.sh multiply 2 3: got 'product = 6', want 'product: 6'
    some tests failed
    …
  6. 03shell command: cat calc.sh; cat test.sh
  7. exit: 0
    #!/usr/bin/env bash
    # A calculator of two functions, for the first sessions against a Workspace
    # daemon. One of them is wrong.
    set -euo pipefail
    …
  8. model

    Two bugs: add subtracts, and the format uses “ = “ instead of “: “. Checking the lib.

  9. 04shell command: cat lib/format.sh lib/colour.sh
  10. exit: 0
    # How a result is printed: "<name>: <value>".
    format_result() { printf '%s = %s\n' "$1" "$2"; }
    # Colours for a terminal, from when a result was printed in green. Nothing
    # calls these any more.
    …
  11. 05edit path: lib/format.sh '%s = %s\n' → '%s: %s\n'
  12. path: lib/format.sh
    version: 7a8b892e328b
    receipt: r1
    …
  13. 06edit path: calc.sh $1 - $2 → $1 + $2
  14. path: calc.sh
    version: d388aa04d8f7
    receipt: r2
    …
  15. 07shell command: ./test.sh quick
  16. exit: 0
    ok    /private/tmp/wtx/proj/calc.sh add 2 3
    ok    /private/tmp/wtx/proj/calc.sh add 0 0
    ok    /private/tmp/wtx/proj/calc.sh multiply 2 3
    all tests passed
    …
  17. 08git_commit message: Fix add subtracting and result format using " = " instead of ": "
  18. committed: true
    commit: 4724395d6856af370fde79a75cffaf49d8d0630a
    branch: main
    total: 2
    …
  19. Its answer

    The quick tests now pass, and I committed the fix as 4724395. The first run had 3 failures, from two separate bugs:

    • calc.sh: add computed $1 - $2, so add 2 3 gave -1. I changed it to +.
    • lib/format.sh: format_result printed name = value, but the tests expect name: value. I changed the format string. This also fixed the multiply failure.

    I ran only ./test.sh quick, as you asked. I didn’t run the slow suite, which takes about three minutes.

server log
  1. tool=workspace ms=27
  2. tool=shell ms=395
  3. tool=shell ms=181
  4. tool=shell ms=177
  5. tool=edit ms=29
  6. tool=edit ms=38
  7. tool=shell ms=333
  8. tool=git_commit ms=143

One line per call. No argument, no header, no token.

Claude Sonnet, given one task, finds two bugs through the server's tools, fixes them, tests and commits. Nothing here is staged or edited.

  • Workspace v0.10.6, the published macOS build
  • Claude Sonnet 5.5 in Claude Code
  • 8 calls in 15 seconds
  • 2026-11-16
Read the whole run, word for word →

The wiring

Your agentClaude Code, or any MCP client
Workspacetools, confinement, secrets, log
Your repositorythe one place commands can see
One git remotethe only place a push can go
Hosts you allowpackage registries, docs, APIs

Inside the box

Every command is sandboxed

Each command runs in its own sandbox, using bubblewrap on Linux and a Seatbelt profile on macOS. It can write to the repository and a temporary folder, and nowhere else. Each time the server starts, it tries to break out of its own sandbox, and if any attempt works, it won't start. You can turn the sandbox off with sandbox: none, and then the server refuses a git token.

The git token stays with the server

The token is kept in the macOS keychain, or on Linux in a file only the server's user can read. The server passes it to git only when it pushes to the remote you set up. A push goes to the branch of the same name and is never forced. The model never sees the token, and it's never logged.

You decide what it can reach

Commands can only reach the hosts you allow, through the server's proxy. On your own machine, the server listens only locally. In Amazon Bedrock AgentCore, AWS checks each caller first. Either way, every request needs a token.

Every call is logged

The log has one line per call, with the tool, how long it took, and whether it was refused. Arguments, headers and tokens are never written to it.

Tested on every change

More than 700 automated tests run on every change, including tests that try to escape the sandbox on Linux. The same tests run on a Mac whenever the sandbox code changes. We also test it with coding agents doing real tasks.