tandem-client

Python client for the Tandem autonomous agent engine HTTP + SSE API.

Install

pip install tandem-client

Python 3.10+ required.

Quick start

import asyncio
from tandem_client import TandemClient

async def main():
    async with TandemClient(
        base_url="http://localhost:39731",
        token="your-engine-token",     # from `tandem-engine token generate`
    ) as client:
        # 1. Create a session
        session_id = await client.sessions.create(
            title="My agent",
            directory="/path/to/my/project",
        )

        # 2. Start an async run
        run = await client.sessions.prompt_async(
            session_id, "Summarize the README and list the top 3 TODOs"
        )

        # 3. Stream the response
        async for event in client.stream(session_id, run.run_id):
            if event.type == "session.response":
                print(event.properties.get("delta", ""), end="", flush=True)
            if event.type in ("run.complete", "run.completed", "run.failed", "session.run.finished"):
                break

asyncio.run(main())

Sync usage (scripts)

from tandem_client import SyncTandemClient

client = SyncTandemClient(base_url="http://localhost:39731", token="...")
session_id = client.sessions.create(title="My agent")
run = client.sessions.prompt_async(session_id, "Analyze this folder")
print(f"Run started: {run.run_id}")
# Note: stream() is async-only; use the async client to receive events
client.close()

API

TandemClient(base_url, token, *, timeout=20.0)

Use as an async context manager or call await client.aclose() manually.

client.sessions

Sampling parameters: temperature, top_p, and max_tokens are optional. Set them on create(...) for a session-level default, and/or on prompt_async(...) to override per prompt (per-prompt wins field by field). Omitting them preserves engine defaults. The engine clamps values to each provider's accepted range (e.g. Anthropic caps temperature at 1.0) and drops temperature for models that reject it (e.g. OpenAI reasoning models) rather than failing the run. Useful for forcing low temperature on strict-JSON roles:

session_id = await client.sessions.create(
    title="reviewer", provider="anthropic", model="claude-sonnet-4-6",
    temperature=0.1, max_tokens=2048,
)
run = await client.sessions.prompt_async(session_id, "...", temperature=0.0)

Prompt with file attachments:

run = await client.sessions.prompt_async_parts(
    session_id,
    [
        {
            "type": "file",
            "mime": "image/png",
            "filename": "diagram.png",
            "url": "/srv/tandem/channel_uploads/telegram/123456789/diagram.png",
        },
        {"type": "text", "text": "Explain this diagram."},
    ],
)

Prompt with browser tools enabled for QA:

run = await client.sessions.prompt_async(
    session_id,
    "Run QA on the target web app. Capture screenshots for failures.",
    tool_mode="required",
    tool_allowlist=[
        "browser_status",
        "browser_open",
        "browser_navigate",
        "browser_snapshot",
        "browser_click",
        "browser_type",
        "browser_press",
        "browser_wait",
        "browser_extract",
        "browser_screenshot",
        "browser_close",
    ],
)

Browser automation via tools

The SDK exposes two different browser paths:

• client.browser for readiness, install, and smoke-test flows
• client.execute_tool(...) or session runs for actual browser automation

The browser namespace does not currently wrap open, click, type, extract, or screenshot. Those actions are exposed as standard engine tools.

status = await client.execute_tool("browser_status", {})

opened = await client.execute_tool("browser_open", {
    "url": "https://example.com",
})
session_id = opened.output["session_id"]

snapshot = await client.execute_tool("browser_snapshot", {
    "session_id": session_id,
    "include_screenshot": True,
})

html = await client.execute_tool("browser_extract", {
    "session_id": session_id,
    "format": "html",
})

await client.execute_tool("browser_close", {
    "session_id": session_id,
})

Use this pattern for agents that need to open pages, click elements, enter text, wait for content, extract page HTML or text, and capture screenshots through the engine.

client.routines

Create a cron routine:

await client.routines.create({
    "name": "Daily digest",
    "schedule": "0 8 * * *",
    "prompt": "Summarize today's activity and write a report to daily-digest.md",
    "allowed_tools": ["read", "write", "websearch"],
})

client.automations_v2

automation = await client.automations_v2.create({
    "name": "Daily Marketing Engine",
    "status": "active",
    "schedule": {
        "type": "interval",
        "interval_seconds": 86400,
        "timezone": "UTC",
        "misfire_policy": "run_once",
    },
    "agents": [
        {
            "agent_id": "research",
            "display_name": "Research",
            "model_policy": {
                "default_model": {
                    "provider_id": "openrouter",
                    "model_id": "openai/gpt-4o-mini",
                }
            },
            "tool_policy": {"allowlist": ["read", "websearch"], "denylist": []},
            "mcp_policy": {"allowed_servers": []},
        }
    ],
    "flow": {
        "nodes": [
            {"node_id": "market-scan", "agent_id": "research", "objective": "Find 3 trend signals."}
        ]
    },
})
run = await client.automations_v2.run_now(automation.automation_id or "")

webhook = await client.automations_v2.create_webhook_trigger(
    automation.automation_id or "",
    {
        "name": "GitHub issues",
        "provider": "github",
        "providerEventKind": "issues.opened",
        "defaultDataClass": "customer_data",
        "defaultRiskTier": "internal_write",
    },
)
print(webhook.trigger.callback_url)
print(webhook.new_secret)  # Returned once; store it securely.

deliveries = await client.automations_v2.list_webhook_deliveries(
    automation.automation_id or "",
    webhook.trigger.trigger_id or "",
)

client.workflow_plans

preview = await client.workflow_plans.preview(
    prompt="Create a release checklist automation",
    plan_source="planner_page",
)

started = await client.workflow_plans.chat_start(
    prompt="Create a release checklist automation",
    plan_source="planner_page",
)

updated = await client.workflow_plans.chat_message(
    plan_id=started.plan.plan_id or "",
    message="Add a smoke-test step before rollout.",
)

applied = await client.workflow_plans.apply(
    plan_id=updated.plan.plan_id,
    creator_id="operator-1",
)

import_preview = await client.workflow_plans.import_preview(
    bundle=applied.plan_package_bundle,
)

imported = await client.workflow_plans.import_plan(
    bundle=import_preview.bundle or applied.plan_package_bundle,
)

Planner page workflow

The Planner page uses the same workflow_plans surface and keeps the plan bundle as the portable artifact. A minimal end-to-end flow looks like this:

started = await client.workflow_plans.chat_start(
    prompt="Plan a release workflow with approval and handoff",
    plan_source="intent_planner_page",
    workspace_root="/workspace/repos/tandem",
)

revised = await client.workflow_plans.chat_message(
    plan_id=started.plan.plan_id or "",
    message="Split the work into review, validate, and publish phases.",
)

applied = await client.workflow_plans.apply(
    plan_id=revised.plan.plan_id or "",
    creator_id="planner-operator",
)

preview_import = await client.workflow_plans.import_preview(
    bundle=applied.plan_package_bundle,
)

if preview_import.import_validation.get("compatible"):
    await client.workflow_plans.import_plan(
        bundle=preview_import.bundle or applied.plan_package_bundle,
    )

Use this flow when you want the same governed bundle the control-panel Planner page hands to Automations, Coding, and Orchestrator.

Additional namespaces

The Python SDK already includes the newer engine surfaces that have landed across the repo:

• client.browser for status(), install(), and smoke_test()
• client.storage for storage file inspection and legacy repair scan helpers
• client.workflows for workflow registry, runs, hooks, simulation, and live events
• client.resources for key-value resources
• client.skills for list/get/import plus preview, templates, validation, routing, evals, compile, and generate flows
• client.packs and client.capabilities for pack lifecycle and capability resolution
• client.automations_v2, client.incident_monitor, client.coder, client.agent_teams, and client.missions for newer orchestration APIs

Long-running orchestrations

draft = await client.orchestrations.create({
    "orchestration_id": "quarter-close",
    "name": "Quarter close",
    "root_node_id": "collect",
    "nodes": [{
        "node_id": "collect",
        "name": "Collect evidence",
        "position": {"x": 0, "y": 0},
        "kind": "workflow",
        "automation_id": "collect-evidence",
        "allowed_transition_keys": ["complete"],
    }, {
        "node_id": "done",
        "name": "Done",
        "position": {"x": 300, "y": 0},
        "kind": "terminal",
        "outcome": "complete",
    }],
    "edges": [{
        "edge_id": "collect-complete",
        "from_node_id": "collect",
        "to_node_id": "done",
        "transition_key": "complete",
    }],
    "goal_policy": {"max_hops": 20, "on_limit": "pause_for_review"},
})
validation = await client.orchestrations.validate("quarter-close")
stale = await client.orchestrations.stale_references("quarter-close")
refreshed = await client.orchestrations.refresh_references(
    "quarter-close", expected_updated_at_ms=draft.updated_at_ms
)
published = await client.orchestrations.publish("quarter-close")
started = await client.stateful_runtime.start_goal(
    orchestration_id="quarter-close",
    orchestration_version=published.orchestration.version,
    objective="Complete the quarter close with verified evidence",
    idempotency_key="quarter-close-2026-q2",
)
events = await client.stateful_runtime.list_goal_events(
    started.goal.goal_id, cursor=0, limit=100
)
async for event in client.stateful_runtime.events(
    started.goal.goal_id, cursor=events.last_cursor
):
    print(event.cursor, event.type, event.properties)

completed = await client.stateful_runtime.settle_goal_completion(
    started.goal.goal_id,
    transition_key="complete",
    final_artifact={"artifact_type": "report", "value": {"ok": True}},
)

Orchestration authoring uses the canonical draft-v0 routes. Draft writes contain only writable graph fields; server-owned fields such as status, version, tenant context, and timestamps are returned by the server. get(id) returns the draft and latest published version together, while get_version(id, version) returns an immutable published snapshot. Goal event cursors are durable and may be passed back to both list_goal_events(..., cursor=...) and events(..., cursor=...) for exact replay.

• For the Incident Monitor flow, see Incident Monitor

browser = await client.browser.status()
storage_files = await client.storage.list_files(path="data/context-runs", limit=100)
workflows = await client.workflows.list()
resources = await client.resources.list(prefix="agent-config/")
templates = await client.skills.templates()

Storage archive cleanup and root JSON migration are local maintenance operations. Run them with the engine CLI, for example tandem-engine storage cleanup --dry-run --context-runs --json.

client.coder

client.coder also includes project-scoped GitHub Project intake helpers:

await client.coder.put_project_binding("repo-123", {
    "github_project_binding": {
        "owner": "acme-inc",
        "project_number": 7,
        "repo_slug": "acme-inc/tandem",
    }
})

inbox = await client.coder.get_project_github_inbox("repo-123")
intake = await client.coder.intake_project_item("repo-123", {
    "project_item_id": inbox.items[0].project_item_id,
    "source_client": "sdk_test",
})

client.agent_teams template management

await client.agent_teams.create_template({"templateID": "marketing-writer", "role": "worker"})
await client.agent_teams.update_template("marketing-writer", {"system_prompt": "Write concise copy."})
await client.agent_teams.delete_template("marketing-writer")

client.mcp

await client.mcp.add(
    "arcade",
    "https://mcp.arcade.ai/mcp",
    allowed_tools=["search", "search_docs"],
)
await client.mcp.connect("arcade")
tools = await client.mcp.list_tools()
await client.mcp.patch("arcade", allowed_tools=["search"])
await client.mcp.patch("arcade", clear_allowed_tools=True)

client.channels

await client.channels.put("discord", {
    "bot_token": "bot:xxx",
    "guild_id": "1234567890",
    "security_profile": "public_demo",
})
status = await client.channels.status()
config = await client.channels.config()
prefs = await client.channels.tool_preferences("discord")
await client.channels.set_tool_preferences("discord", {"disabled_tools": ["webfetch_html"]})
verification = await client.channels.verify("discord")
print(status.discord.connected)
print(config.discord.security_profile)
print(prefs.enabled_tools)
print(verification.ok)

client.packs

packs = await client.packs.list()
detected = await client.packs.detect(path="/tmp/my-pack.zip")
if detected.get("is_pack"):
    await client.packs.install(path="/tmp/my-pack.zip", source={"kind": "local"})

client.capabilities

bindings = await client.capabilities.get_bindings()
discovery = await client.capabilities.discovery()
resolution = await client.capabilities.resolve(
    {
        "workflow_id": "wf-pr",
        "required_capabilities": ["github.create_pull_request"],
    }
)

client.permissions

snapshot = await client.permissions.list()
for req in snapshot.requests:
    await client.permissions.reply(req.id, "allow")

client.memory

# Put (SDK accepts `text`; server persists global `content`)
await client.memory.put(
    "Use WAL mode for sqlite in long-lived services.",
    run_id="run-123",
)

# Search
result = await client.memory.search("sqlite wal", limit=5)

# List by user scope
listing = await client.memory.list(user_id="user-123", q="sqlite")

# Audit
audit = await client.memory.audit(run_id="run-123")

# Promote / demote / delete
await client.memory.promote(listing.items[0].id)
await client.memory.demote(listing.items[0].id, run_id="run-123")
await client.memory.delete(listing.items[0].id)

client.providers

catalog = await client.providers.catalog()
await client.providers.set_defaults("openrouter", "anthropic/claude-3.7-sonnet")
await client.providers.set_api_key("openrouter", "sk-or-...")

Common event types

License

MIT