Share Results With a Team

Once one person has computed something, nobody else has to compute it again. You run one barca serve that everyone points at, and you keep using barca get exactly as before.

Only need to run on the server?

Start barca serve --host 0.0.0.0 in the deployed project, then use these commands from your laptop (replace the address with your server):

barca get revenue --server http://barca.internal:8274
barca run publish --server http://barca.internal:8274
barca history --server http://barca.internal:8274

This executes the server’s copy of the project and keeps results and history there. It needs no bucket on your laptop and does not upload local edits. Keep the server on a private network or behind an authenticating proxy; it has no login.

Also want laptops to compute against the same cache?

Follow the three steps below. With this setup, plain barca get computes on the laptop and records the run on the server; barca get --server computes on the server. Both use one shared record of results and runs. --local deliberately uses a separate local record.

You need:

a machine your team can reach, such as a small VM or a box on your VPN or Tailscale
a bucket on S3, GCS or Azure that the server and every laptop can read and write
the project in git, so the server and the laptops have the same code

1. Add two sections to `barca.toml`

# barca.toml, committed with the project
[server]
url = "http://barca.internal:8274"   # where step 2 runs

[remote]
uri = "s3://team-bucket/barca"       # gs:// and az:// work too

Commit it. Everyone, the server included, uses this same file. barca serve ignores the [server] section, and each laptop reads it to find the server.

Install the extra for your bucket, for example with uv add 'barca[s3]' (or barca[gcs] or barca[azure]). Bucket credentials come from the usual places, such as AWS_PROFILE, so barca docs remote has the details.

2. Start the server

On the server, in a checkout of the project:

uv sync
uv run barca serve --host 0.0.0.0

...
[barca] serving on http://0.0.0.0:8274  (1 file)
[barca] warning: listening on 0.0.0.0 with no authentication — anyone who can reach
port 8274 can trigger runs

The web UI is at http://barca.internal:8274/. There is no login, so keep the server on a private network or behind a proxy that authenticates (Deploying).

3. Use barca as usual

On your laptop nothing changes. The first barca get computes what is new and records it on the server:

$ barca get revenue
[barca] server http://barca.internal:8274 ([server] url in barca.toml)
[barca] 2/2 steps | done in 2.1s
...
Run 90af4d65f284 | got 'revenue' in 2.068s (2 steps, 1 phase)

.barca/artifacts/pipeline.py--revenue/43fbf60a1f6e...json
  {
    "total": 42
  }

Your teammate runs the same command in their own checkout. Nothing runs: both results come from the server’s record, and the one they asked for is downloaded from the bucket:

$ barca get revenue
[barca] server http://barca.internal:8274 ([server] url in barca.toml)
[barca] checking artifact store s3://team-bucket/barca/default/artifacts (timeout 10s)
[barca] fetched 1 cached artifact (13 B) in 0.0s
Run 9025447e887c | got 'revenue' in 0.048s (0 steps, 1 phase)

.barca/artifacts/pipeline.py--revenue/43fbf60a1f6e...json
  {
    "total": 42
  }

If they have edited revenue, only revenue runs on their machine. orders still comes from the cache, because a result is shared whenever the code and inputs match.

What else you can do

barca status                 # what the team has, and each result's shape
barca history                # everyone's runs
barca history --status failed --since 24h   # what failed today, each with its step and error
barca history 836fdca186c8   # one run: its steps, the error and traceback, and the logs command
barca sql "select * from revenue"   # query the team's results here (the server runs no SQL)
barca get revenue --server   # run it on the server
barca get revenue --local    # skip the server once

$ barca history -l 3
RUN_ID         CMD     TARGET                   STATUS      STEPS CACHED   TIME STARTED
------------------------------------------------------------------------------------------------------
9086ad528e03   get     revenue                  success         0      2   0.0s 2026-10-22 19:00:53 UTC
9085f266d644   get     revenue                  success         2      0   2.1s 2026-10-22 19:00:46 UTC

$ barca history --status failed
RUN_ID         CMD     TARGET                   STATUS      STEPS CACHED   TIME STARTED
------------------------------------------------------------------------------------------------------
836f80e59467   get     enrich                   failed          1      1   0.1s 2026-10-22 22:57:27 UTC  enrich: ConnectionError: pricing API timed out
836fdca186c8   get     clean                    failed          1      1   0.1s 2026-10-22 22:57:27 UTC  clean: ZeroDivisionError: division by zero

--server runs the server’s copy of the code, not your local edits, and prints the output here. Use it for heavy jobs, or for anything that must run where the data lives. From Python, barca.get("revenue") follows the same barca.toml, and barca.get("revenue", server=True) runs it on the server.

Schedules run on the server. Give an asset freshness=Schedule("0 6 * * *") and the server brings it up to date every day at 06:00 (Scheduling).

If something goes wrong

You see	It means	Do
`the barca server at ... did not answer GET /health` (exit 3)	the server is down or unreachable	start it, or add `--local` to work alone for now
`a server is configured (...) but this project has no artifact store` (exit 2)	the laptop has `[server]` but no `[remote]`	add the `[remote]` section from step 1
`the barca server at ... writes artifacts to ..., this project to ...` (exit 2)	the server and the laptop use different buckets	use the one committed `barca.toml`
`warning: run ... was not recorded on the server`	your run worked, but the server did not get it	nothing: the next command sends it again

Every message ends with the fix, and --local is always offered. barca never quietly falls back to your own cache when the server is down, because two people who stop agreeing without noticing is what the server is there to prevent. --local says you mean it.

Deploying: run the server under systemd or Docker, behind nginx or Traefik.
Server Mode reference: every rule, including what is shared, what happens when a post fails, and the HTTP routes. In the terminal, run barca docs server-mode.