For teams running AI agents in production. See every run and what it cost, get an incident when work goes missing, stop a run at its budget, and put a person in front of the actions that matter.
Daily digest didn't report by 09:00SK Sent to #ops · Sam is looking
Stopped at budget$5.00 of $5.00 · Nightly research
Illustration with sample data
Follow one agent run from its first report to the review
Five things Toolcaise does, shown on sample runs. Each step says which way of connecting it needs, because not every connection can do everything.
Each run reports what it did and what it cost
Steps, model and tool calls, tokens and cost arrive as the agent works. Open any run later and see how it went.
Works with the Python and Node.js SDKs, OpenTelemetry, the HTTP API or Toolcaise Connect. Each one reports what it can see.
Runs · Nightly research
Nightly research, run 184Started 02:00 on piaiRunning
ModelPlan the research2,410 tokens
Toolsearch_web6 calls
Toolfetch_page14 calls
ModelSummarise sources9,820 tokens
ModelWrite the brief6,190 tokens
Tokens18,420Cost$1.84Tool calls20
Illustration with sample data
A supervised run stops at its budget
Wrap the command in Toolcaise Connect and give it a cost, token, call or runtime limit. At the limit, Connect ends the command and everything it started, even with no connection to Toolcaise.
Needs Toolcaise Connect on the machine that runs the agent, and a price list you supply for cost limits. Through the SDKs, stop and cancel are cooperative: your code has to check for them.
Limit reached. The command and the 3 processes it started were stopped.Exit code 124 · run recorded as canceled at its cost limit
Example command and sample output
A person approves the action before it happens
Before a sensitive step the agent waits for someone on your team. The approval covers that one call: the same tool with different arguments needs a new decision.
Works through an SDK checkpoint in your code, or Connect’s MCP gateway (Connect 0.3 or later). For MCP calls the approver sees the tool name, not the arguments.
Approvals
GitHub tools is waitingApprove a tool call?Held
github__merge_pull_request
The arguments stay on your machine. This approval covers this one call only.
MRMaya R.Engineering lead
Deny Approve
Approved by Maya R. The call ran once.
Illustration with sample data
A missed schedule turns into an incident
Tell Toolcaise when an agent should run. If it doesn’t report in time, stalls partway or keeps failing, an incident opens and your team hears about it.
Works with every way of connecting. Alerts go to Slack, Microsoft Teams, Discord, PagerDuty or a webhook. Checks run about once a minute, so an alert isn’t instant.
Incidents · Daily digest
Expected by 09:00 on weekdays
Mon08:12
Tue08:19
Wed08:26
Thu08:33
FriNo report
Daily digest didn't report by 09:00Incident opened at 09:00 · Missed schedule
##opsToolcaise: Daily digest didn't report by 09:00. Open the incident.SK
Illustration with sample data
Someone checks what the agent produced
A run that finished isn’t always work that’s right. A reviewer opens the result and accepts or rejects it.
Toolcaise records the decision. It can’t check on its own that the work is correct, which is why this step is a person.
Runs · Month-end reconciliation
The agent reported
CompletedReconciled 42 invoices against the ledger
Outcome review
PAPriya A.FinanceRejected
Reviewed at 10:42
The agent's claim and the review are recorded separately, so neither is mistaken for the other.
Illustration with sample data
Your agents keep running where they run today
Nothing moves into Toolcaise. Agents report over outbound HTTPS, so there is no inbound port to open, and your model-provider keys stay with you. Each agent needs a short setup step: Toolcaise doesn’t find agents on its own.
What each way of connecting can see, and what it can stop
Seeing a run and being able to stop it are different things. Pick the connection that gives you the control you need.
Cooperative
Python and Node.js SDKs, HTTP
What Toolcaise sees
What the agent reports: runs, steps, usage, outcomes
What it can stop or hold
Pause, resume, cancel and exact-action approvals, at checkpoints your code implements. Cooperative: code that skips the checkpoint is not stopped.
Observes only
OpenTelemetry
What Toolcaise sees
Runs, model and tool spans, and token counts from your framework’s own tracing
What it can stop or hold
Nothing. It observes only.
Observes only
Connect: Claude Code and Codex adapters
What Toolcaise sees
The allow-listed events and metrics those tools emit
What it can stop or hold
Nothing. It reports telemetry only.
Enforced on your machine
Connect: supervised run
What Toolcaise sees
Every model call made through an OpenAI- or Anthropic-compatible API, measured on your machine, plus the process’s start, end and resources
What it can stop or hold
Runtime, cost, token and model-call limits and a model allowlist, by stopping the whole process tree. Pause, resume, stop and limits from the dashboard.
Enforced at the proxy
Connect: MCP tool gateway
What Toolcaise sees
Every tool call that passes through the proxy
What it can stop or hold
Allow or deny by tool name, a call budget per session, and a hold until a person approves the exact call (Connect 0.3 or later). The approver sees the tool name; the arguments stay on your machine.
What no connection can do
Tell whether an agent’s work is correct. A person reviews that.
Stop an agent that goes around every control point.
Measure model calls sent to an endpoint Connect doesn’t sit in front of. The runtime limit and the stop still end the run.
Every plan includes monitoring, incidents and alerts, approvals and Toolcaise Connect limits. Plans differ in how many agents and people a workspace holds.
It is the operating routine around agent work: observing activity, investigating exceptions, reviewing outputs, and deciding when an integrated action may proceed. Toolcaise brings these signals and decisions into a shared workspace.
Can I keep the agents and infrastructure I already use?
Yes. Your agent keeps running where it runs today. Connect it with the Python or Node.js SDK, the HTTP contract, the OpenTelemetry traces your framework already emits, or Toolcaise Connect on the machine that runs it. Toolcaise does not discover agents on its own: each path needs a short setup step.
How is this different from an agent framework?
An agent framework helps define and execute the workflow. Toolcaise provides a supervision layer around workflows you operate: reported run history, incidents, human outcome review, and compatible approval checkpoints.
Will Toolcaise stop an agent that runs away?
It stops what it sits in front of. A run wrapped in Toolcaise Connect is stopped, with every process it started, at the cost, token, call or runtime limit you set, and can be paused or stopped from the dashboard. Through the SDKs, pause and cancel are cooperative and apply at checkpoints your code implements. OpenTelemetry only observes. None of them can judge whether an action is safe for your business; that is what approvals and review are for.
Do the images or example counts show real customer results?
The product illustrations and diagrams use labeled sample data, not live workspace screenshots. They do not represent customer performance or independent verification of agent execution.
What does it cost?
Free covers 3 agents and 2 members. Team is $49 a month for up to 25 agents, and Business is $199 a month for up to 100. Paid plans are billed monthly through Stripe and can be cancelled from the billing portal. For more agents or a pilot on specific terms, talk to us.
Connect your first agent
Create a free workspace, connect the agent you already run, and watch its next run arrive. Free covers 3 agents, with no card required.