Wisp turns outside signals into proposals you review.
Wisp checks events from connected feeds against your sites, suppliers and open orders. What touches your operation becomes a proposal with its evidence, and it waits until a person approves it.
One agent for each job
Agents work within the permissions of the person they run for. They prepare findings, drafts and proposals, and people review them.
- LiveHeartbeatOn a schedule or an eventRuns the checks you schedule and puts its findings on the map and in Wisp.
- PreviewResearchWhen you askResearches a site, region or selection with cited sources and adds the result to a dossier for review.
- LiveGuided interviewDuring a captureAsks a contributor one question at a time about their work, then drafts the BPMN for them to check.
- LiveDashboard builderWhen you askBuilds and edits dashboards from your process data and explains a chart on request.
- LiveStewardEvery chat messageReads each message first, answers what it can and hands a short brief to the right specialist.
- LiveYour own skillsWritten by your adminsAdmins write skills for your processes and assign them to the agent roles that should use them.
You decide what an agent may change
- Review
AI changes arrive as a batch you keep or undo
In Review mode Wisp applies the changes from each chat turn as one batch and lists them on a review card. You keep or undo the batch as a whole.
- Admins choose Read, Review or Write for AI modeling
- Undo restores the earlier state, and every decision is saved
- Govern
Agents act with the rights of the person behind them
A chat run uses the permissions of the person who started it. Scheduled work runs for a named sponsor, and each operation is checked again before it runs.
- Revoking a person's access stops the work that depends on it
- The person who started a run can stop it, and org admins can stop any scheduled check
- Audit
Every run leaves a record
Admins see each AI run with its agents, tool calls, usage and cost, and can export it. In chat, each answer lists the agents that worked on it.
- Chat turns are stored step by step, so a long thread can be traced
- Export and retention are set per organization
How the controls work
- The agent harnessTool search, skills, compaction, parallel specialists, goal runs and the sandbox, shown step by step.
- Keep or undo AI changesHow AI edits arrive in batches that you review in Blueprint.
- Delegation and permission checksAgents run with the requester's rights or a sponsor's, checked on every operation.
- Turn records and the audit logEach AI run and chat turn is stored with its agents, tools and cost.
- GeistBenchOur public benchmark for AI agents on enterprise process tasks. Results appear after review.
What Wisp and the agents do today
The current feature set, grouped by job.
Agents
- AI Team with named teammates in Wisp
- Heartbeat checks on a schedule or triggered by events
- Goal mode for longer tasks, with budgets per round
- Skills that admins write and assign to agent roles
Control
- Read, Review or Write permission for AI modeling
- Keep or undo per batch, one batch per AI turn
- Stop per run, plus an emergency pause for admins
- Spend limits per provider, organization and user
Models and keys
- Three tiers: pro, standard and fast, with the provider chosen per tier
- Your own provider keys, stored encrypted per organization
- Model preferences per agent
- AI spend view with cost per provider and per product
Use your own keys or ours
Admins pick the provider for each tier, and the model catalog picks the model. Keys you bring are stored encrypted per organization and can be tested, rotated and revoked in Connections & Keys.
- Anthropic
- OpenAI
- Gemini
- Mistral
- DeepSeek
- Qwen
- Meta
- Kimi
- Z.ai
- Grok
Put an agent on one of your processes
Start with a scheduled check or a guided interview on a process your team runs today.