fountain-agent-sdk

Give an agent a computer, your repositories, and your credentials in one call.

from fountain import Fountain

fountain = Fountain()
result = fountain.run(
    "Upgrade us to Phoenix 1.8 and open a PR",
    agent="reposage",
    vault="github-bot",
).result()

print(result.text)
print(result.url)

The agent runs on a real sandbox with the selected environment and vault. The sandbox remains after the turn, so a follow-up continues on the same computer and in the same agent session.

Install

pip install fountain-agent-sdk

Python 3.9 or newer. The SDK has no runtime dependencies.

Credentials

Fountain() resolves credentials the same way as the Fountain CLI:

api_key:  argument -> FOUNTAIN_API_KEY -> FOUNTAIN_TOKEN -> ~/.fountain/credentials
base_url: argument -> FOUNTAIN_BASE_URL -> ~/.fountain/credentials -> hosted Fountain

Use profile="work" to select another credentials-file profile. A client in a Fountain sandbox automatically uses its conversation-scoped token and marks the work it creates as child conversations.

Wait, stream, or fan out

run() starts work immediately in a background thread. Calling result() waits for the finished turn. Iterating the same handle streams its events.

run = fountain.run("Review this repository", agent="reviewer")

for event in run:
    if event["type"] == "tool":
        print("->", event["name"])
    elif event["type"] == "text":
        print(event["text"], end="", flush=True)

result = run.result()  # this is the same run; no second request is made

For only the answer text:

run = fountain.run("Review this repository", agent="reviewer")
for chunk in run.text_stream:
    print(chunk, end="", flush=True)

The same handle also works in asyncio code. HTTP still runs in the SDK's background thread, so it does not block the event loop.

result = await fountain.run("Review this repository", agent="reviewer")

run = fountain.run("Review another repository", agent="reviewer")
async for event in run:
    print(event)

Start several handles before collecting them to provision their sandboxes in parallel:

runs = [fountain.run(prompt, agent=agent) for agent in agents]
results = [run.result() for run in runs]

A failed agent turn is a RunResult with state == "failed". Rejected HTTP requests, transport failures, and SDK wait timeouts raise typed exceptions.

Follow-ups and permissions

first = fountain.run("Find every N+1 query", agent="reposage").result()
second = fountain.resume(first.conversation_id).send("Fix the worst three.").result()

An agent with an ask permission policy emits permission events. Answer with one of the option ids it offered:

run = fountain.run("Clean the build tree", agent="reposage")
for event in run:
    if event["type"] != "permission":
        continue
    request = event["request"]
    allow = next(option for option in request["options"] if option.get("kind") == "allow_once")
    run.answer(request["request_id"], allow["option_id"])

Resources and teammates

Resource definitions use the API's snake_case keys, so the same dictionary can also appear in a fountain.yml manifest or raw REST request.

environment = fountain.environments.create({
    "name": "fountain-ci",
    "packages": {"apt": ["ripgrep"]},
    "repositories": [{
        "url": "https://github.com/BinaryBourbon/fountain",
        "mount_path": "/work/fountain",
    }],
})

vault = fountain.vaults.create({"name": "github-bot"})
fountain.vaults.secrets.set("github-bot", "GITHUB_TOKEN", token)

agent = fountain.agents.create({
    "name": "reposage",
    "runtime": "claude",
    "model": "anthropic/claude-sonnet-5",
    "environment_id": environment["id"],
    "allowed_vault_ids": [vault["id"]],
})

Agents, environments, and vaults each have list, get, create, update, and delete. Environments and vaults also have write-only secrets helpers.

fountain.team.add("watchtower", name="Watchtower")
reply = fountain.team.message("watchtower", "Any disks over 80%?").result()
fountain.team.schedules.create("watchtower", {
    "cron": "0 9 * * *",
    "prompt": "Check disk usage and say only what changed.",
})

Cancel a wait

run.cancel() stops the local wait and skips the final status request. The agent continues its work. A silent stream can delay cancellation by up to five seconds. Idle reads reconnect from the last complete event.

A custom HTTP transport must honor the supplied socket timeout.

Raw API and errors

Every endpoint remains available through the authenticated HTTP layer:

rows = fountain.request("GET", "/api/audit", query={"limit": 50})

Catch subclasses such as ConversationBusyError, NotReadyError, QuotaExceededError, ValidationError, and AuthError. Every FountainError carries status, code, body, retry_after, and a retryable property.

Develop

python3 -m unittest discover -s tests -v
python3 -m pip wheel . --no-deps