Agents

MindGraph agents are long-running knowledge workers with durable memory, scoped context, explicit tool access, and an audit trail. They can research, maintain Project knowledge, use approved MCP tools, and return for scheduled work without starting from zero.

Configure and supervise agents in the dashboard. The HTTP surface below powers the same sessions, schedules, scope controls, and activity views.

Templates

Templates provide role-specific instructions, a current default model, starter prompts, and suggested integrations. You can change the instructions, model, budget, Project scope, and connected tools when you create an agent.

Create and manage agents

curl -X POST https://api.mindgraph.cloud/v1/agents \
  -H "Authorization: Bearer mg_..." \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Account Intelligence",
    "template_type": "client_research",
    "custom_instructions": "Maintain a cited dossier and flag material changes.",
    "model": "claude-opus-4-6",
    "daily_credit_budget": 500
  }'

Scope and identity

The server resolves agent identity from its bound credential and re-evaluates access for each request. The agent cannot expand its own authority by supplying another agent_id.

• Project scope — set a default Project on the agent, then narrow or override it for a session or schedule where permitted.
• Space grants — grant read or write access and, when needed, a classification ceiling.
• Project roles — use viewer, editor, or agent roles; corpus admission alone is not authorization.
• Provenance — graph versions and execution records identify the agent and delegated request context.

A broad organization-wide context is available, but Project scope is the safer default for work that should stay inside a known source boundary. See Governance & Access.

Tool grants and MCP

Connecting a service does not automatically expose all of its tools to every agent. The connection is inspected and synchronized into a registry, then tools are granted to an agent explicitly. Native and MCP tools pass through the same authorization and policy checks.

Owners and admins configure these grants in the agent's Governance tab. See MCP Connections for connection custody, registry synchronization, and approval modes.

Policies and approvals

Before a consequential tool call runs, MindGraph checks identity, context scope, tool grant, and published governance policy. A policy can allow, deny, or require human approval. Approval-required calls pause with their arguments and resume only after an authorized dashboard user decides; access and registry drift are checked again before execution.

Policy decisions, approvals, executions, and resulting graph changes remain available in the governance audit trail. Revoking a grant or pausing an agent applies to the next request.

Sessions

Schedules

A schedule combines a cron expression, message template, budget, and optional Project scope. Use it for recurring briefs, account monitoring, knowledge maintenance, or status reviews.

Activity and audit