Connect an agent
Start with the activity and run history of the agents you already run. Then add outcome review, and approval checkpoints where your team needs to decide.
Create a credential
In the dashboard, under Agents, name your agent and copy its credential. It is shown once, belongs to that one agent, and can be rotated or revoked at any time.
Choose how it reports
Wrap your code with the Python or Node.js SDK, point an OpenTelemetry exporter at Toolcaise, or send HTTP requests from a workflow tool. Nothing on your side needs an inbound port or a VPN, and your model-provider keys stay with your agent.
See the real work
The setup screen confirms the first report. Runs then show their steps, model calls, token counts and outcomes, and incidents open when a heartbeat or a schedule is missed.
Choose your agent environment
Python 3.10+, standard library only. Download the SDK beside your agent, check the credential, then wrap the work your agent already does.
Invoke-WebRequest https://toolcaise.com/downloads/toolcaise_agent.py -OutFile toolcaise_agent.py
$secret = Read-Host 'Agent credential' -AsSecureString
$env:TOOLCAISE_AGENT_TOKEN = [System.Net.NetworkCredential]::new('', $secret).Password
python toolcaise_agent.py checkConnect your real work
from toolcaise_agent import Toolcaise
from openai import OpenAI
from your_agent import main # your existing function
client = Toolcaise()
# Optional: record each model call with its model, token counts and latency.
# Prompts and output are never sent. instrument_anthropic() does the same for Anthropic.
openai = client.instrument_openai(OpenAI())
with client.run("daily-work"):
with client.span("agent-work"):
main(openai)
client.outcome("daily report", status="succeeded")
# Add client.checkpoint() between safe steps to let a person pause or cancel the run.Replace the example imports and image names with your own. The SDK sends metadata only, never prompts, output or exception text. Runs send a heartbeat every 60 seconds. Reporting never raises and never waits on the network: if Toolcaise is unreachable, reports are queued and retried, and your agent carries on. Checkpoints and guarded actions do the opposite and refuse to proceed when Toolcaise cannot confirm a decision. SDK reference · Framework quickstarts
If Toolcaise is unreachable, your agent keeps running and guarded actions stay blocked
The SDKs queue reports in memory and send them in the background. If Toolcaise is slow or unreachable, your agent carries on and the reports follow when it is back. Checkpoints and guarded actions work the other way: when Toolcaise cannot confirm a decision, the guarded action does not run.
How delivery and retries work →Checkpoints and approvals →Usage, output checks and recovery evidence
The same credential reports per-call context and token use, cumulative run costs, process resources, and checks of the files a run produced, without sending the files. Thresholds on them are set under Supervision in the dashboard, where recovery steps, impacts and handoffs are reviewed too. A person reviews the evidence and authorizes an exact recovery step before the agent may claim it.
Built on a framework? Start with its quickstart
Each quickstart names the path that works best for that framework and exactly what it reports, including what it cannot see.
LangChain and LangGraph
OpenInference traces: every model call with token counts, tool calls and graph steps.
ExploreCrewAI
Your crew's agents, tasks and tools, and its model calls through the provider's instrumentation.
ExploreOpenAI Agents SDK
An instrumented client records every model call; checkpoints and approvals work inside your tools.
ExploreClaude Agent SDK
Each session as a run with its usage and cost, and a person's approval before sensitive tools run.
ExploreVercel AI SDK
AI SDK 7's OpenTelemetry spans: every generation step with token counts, and every tool call.
ExploreHTTP API
Every endpoint, field and limit, from the same OpenAPI document the API is tested against.
ExploreSupervise an agent from outside it
Toolcaise Connect runs on your machine. It can start an agent behind a local gateway that measures every model call itself and stops the run at a time, token, cost or call limit, even with no network path to Toolcaise. It can also gate an MCP server's tools, and turn on the Claude Code and Codex adapters.
The current release is 0.4.0-preview.2, and its binaries are not yet code-signed.
What the record proves, and what it doesn’t
Toolcaise records what your agent reports. A completed run or a sent receipt is the agent's claim, not independently verified execution. Toolcaise Connect's supervised runs measure model calls on your machine rather than taking the agent's word for them. Creating, rotating and revoking a credential are recorded as Toolcaise events, and revoking one stops new reports while keeping the history.
Send only operational metadata you intend to store. The SDKs never send prompts, output or exception text; fields named like secrets are refused, but free text you write, such as an event message, is stored as written.
Security and product boundaries →