Understand how to connect your agents, control their actions and investigate what happened—with evidence you can verify.
HaltState checks connected agent actions against your policies before they run. It helps you allow permitted work, block prohibited actions, route decisions for review and investigate the resulting evidence. Place the check at the actual tool or business-action boundary so an action cannot proceed through that integration without a decision.
Yes. Start with a HaltState account and connect an agent using the SDK for your application: Python, TypeScript/JavaScript, Go, Java or Rust. Begin with one meaningful action, such as issuing a refund or sending a customer message, and follow that SDK's installation and integration guide.
SDK guides: Python · TypeScript/JavaScript · Go · Java · Rust. Choose the documented release for the features you need.
No. HaltState is designed to govern connected business actions, including customer communications, refunds and changes to business systems. OpenClaw is one use case; other agents connect at their tool or action boundary using the appropriate integration. A framework name alone does not establish that every tool in that framework is covered.
Define which agent may perform which action, against which resources, and under what conditions. Identify actions that should be allowed, denied or reviewed, then test representative permitted and prohibited requests before relying on the integration. A written policy description is not itself an enforced control: the integration must supply the relevant information and the rule must evaluate it.
Read the governance guide or explore a refund-policy example.
Policy comes first. Requests eligible for council review can receive a council decision; unresolved reviews go to a human. A policy that explicitly requires human approval still requires a human—the council does not override that requirement, and a pending review is not permission to act.
Yes. Select the registered agent, give a reason and choose a pause duration. The stop blocks that agent's subsequent governed checks through the connected enforcement paths. It does not undo a message or refund already sent, guarantee termination of a remote process, or revoke an already issued permit.
In the SDK workflow, your application carries out the action after permission and reports its outcome to HaltState. Investigation separates the policy decision, reported outcome, any recorded external confirmation and evidence integrity. A reported success and a destination-confirmed result are different facts; the record shows which evidence is available.
Yes. Newly created HaltState Proof Packs require a digital signature, with content hashes and verification checks that help detect changes to the recorded evidence. Signing covers the evidence that was recorded; it does not, on its own, prove external delivery or settlement. The synthetic refund demonstration now creates separate, linked signed decision and execution-outcome packs.
Inspect an existing verified approval example or read how signature verification works. That example preserves its original assurance; older evidence is not retrospectively relabelled.
An investigation brings together activity records, decisions, reports, retained content and signed Proof Packs. An activity record is not automatically a signed Proof Pack; the integrity view identifies the evidence available for that record. Older evidence keeps its original verification status rather than being relabelled.
Search the investigation page by agent, customer or conversation reference, time, outcome, recorded message text or external reference. Open the activity to see its linked decision, outcome and available evidence. You can also ask the investigation chat a question such as “What did my support agent send today?”
Searchable content and references must first have been captured and remain within retention and access permissions. Missing content and unresolved links remain visible rather than being filled in by the assistant. Start an investigation.
Yes. Connect a compatible MCP client to HaltState and authorise access to your account's evidence. The connector supports evidence investigation, including search, detail and verification, subject to the permissions granted. It is an investigation interface, not authority for the assistant to change policies or execute business actions.
HaltState records supported connected activity, including policy decisions, action reports and observations that do not request a policy decision. Connected message content can be retained for investigation. It does not automatically see unrelated traffic or tools outside the integration; an observation records activity but does not grant permission.
Yes. New-account capture defaults to enabled with 90-day content retention, with 30-, 90- and 365-day choices and account or per-agent exclusions. Existing explicit settings are preserved. Turning off content retention stops new content capture without disabling governance; retained content is encrypted and kept separately from permanent audit metadata.
Changing capture settings applies to new content; it does not immediately erase content already retained. Read about retention and access controls.
Use Get Started to review the current plans and begin setup. For help inside the product, use support chat; an issue needing follow-up can become a support ticket and receive an email update. Ordinary conversation does not need a ticket, and sending an email to support does not trigger an automatic reply.
Connect a refund, customer message or business-system change. Define its policy, test the decision paths and inspect the evidence.