# How can I control tool access and spending for each AI agent? Give each agent the permissions and optional limits its task needs. By [Agentik](/about.md). Updated: 2026-10-19. A prepaid balance controls funding, while an execution key controls what an agent is authorized to do. Agentik keys can carry supported permissions, tool restrictions and optional spending caps. Use separate keys for different jobs rather than handing every agent the same credential. Limits are opt-in: creating a key does not imply a daily cap has been configured. - [Connect Claude Code](/connect/claude-code.md) ## Separate a research agent from a creative agent Consider a research agent that reads search and SEO data, and a creative agent that generates media. Both can use the workspace balance, but their keys can allow different tools and carry different caps. Choose those controls from the supported key settings and check their saved values. A prompt saying do not exceed the budget is an instruction to the agent. An enforced key cap is a backend control. Use both where appropriate, and keep the human approval required for the requested task. Access controls to configure, not preconfigured limits | Control | Purpose | | --- | --- | | Separate key | Authorize and review one agent's access without sharing its credential with other jobs. | | Allowed tools | Restrict the operations or supported provider patterns the key can use. | | Hourly, daily or lifetime cap | Bound spending in the configured period where supported by the tool's pricing contract. | | Per-tool cap | Limit a particular tool on that execution key. | ## Check whether the tool supports a bounded call A spending cap depends on an enforceable reservation. Tools billed at actual cost without a guaranteed maximum cannot be used through a capped key. Do not remove a cap just to make a blocked operation run; inspect the contract and choose an acceptable operation or policy. A provider rate limit serves a different purpose. It controls request throughput, not your task's total cost. Read the actual error and retry guidance before deciding whether to retry. ## Verify the configuration without an unnecessary purchase Review the saved permissions and limits in the workspace, then let the agent discover and inspect suitable tools. Those read-only steps do not prove that a paid operation would succeed, but they avoid spending just to check a connection. For an authorized production task, retain the run ID and inspect the resulting receipt. Rotate or revoke the relevant credential when its job ends, while preserving the evidence needed to review prior spending. ## What you pay for Supported paid tools use Agentik's prepaid workspace balance. Inspect the selected tool's pricing units and maximum before approving a run; a reservation can differ from its settled charge. Funding fees, taxes and your model's usage are separate. Current terms are on the Pricing and Billing pages linked below. ## Give two agents different authority For a workspace with research and creative agents, configure two execution keys. The research key can be restricted to selected data operations while the creative key can use selected generation tools. Choose optional caps deliberately and verify the saved settings. The table is a proposed permission design, not a statement that these controls are already configured. A tool without an enforceable maximum may not work with a capped key. Give two agents different authority | Agent | Allowed operations | Review | | --- | --- | --- | | Research | Only the selected search and page-reading tools. | Saved allowlist and suitable spending cap. | | Creative | Only the approved generation tools. | Bounded reservation and task-specific approval. | | Offboarding | Revoke the job’s execution key. | Confirm revocation; preserve prior usage records. | ## A prompt to try [Connect Claude Code](/connect/claude-code.md), then use this prompt with your connected assistant. Keep credentials in the authorized connection. ```text Prepare an Agentik access policy for [agents and their jobs]. Identify the exact allowed tools, propose separate execution keys and optional hourly, daily or lifetime caps appropriate to each job. Check which selected tools support bounded reservations. Distinguish prompt instructions from backend controls. Give me a configuration and verification checklist; do not create keys, disclose credentials or loosen an existing limit without approval. ``` ## A related workflow [Keyword research](/keyword-research.md). Seed terms in, a ranked keyword list out, with what page one covers and how people ask in forums. Example tool-call maximum: at most $0.54. 3 seed keywords for the United States in English, 100 ideas scored, page one read for the top 5 keywords, 5 forum searches. This catalog-based reservation maximum belongs to the playbook’s inputs, separately from this guide’s worked example. It is not an executed charge. Optional calls, funding fees, taxes and model usage are excluded. - [dataforseo/dataforseo-labs-google-keyword-ideas](/providers/dataforseo/dataforseo-labs-google-keyword-ideas.md) - [dataforseo/dataforseo-labs-google-bulk-keyword-difficulty](/providers/dataforseo/dataforseo-labs-google-bulk-keyword-difficulty.md) - [dataforseo/dataforseo-labs-google-search-intent](/providers/dataforseo/dataforseo-labs-google-search-intent.md) [View every step, input and tool in the playbook](/keyword-research.md). ## Questions ### How can I set a daily spending limit for each AI agent? Give each agent a dedicated key and explicitly configure its supported daily cap. Confirm the saved setting; keys do not receive an optional cap automatically. ### Can agents share credit with different permissions? Yes. Separate keys can use the same workspace balance while carrying different supported permissions, tool restrictions and spending caps. ### Can a prompt alone enforce a spending limit? A prompt communicates the instruction. Use the supported backend key limits when you need an enforced control, and check that the selected tool supports a bounded reservation. ### Why was an actual-cost tool blocked on my capped key? A tool without a guaranteed maximum cannot honor the same bounded-spend contract. Inspect its pricing and consider a supported bounded alternative rather than bypassing the limit. ## Sources and next steps Agentik publishes this guide. The editorial update date is 2026-10-19; provider terms and availability can change. - [Key limits](/docs/limits.md) - [Pricing and caps](/pricing.md) - [Billing contracts](/docs/billing.md) ### Continue with Agentik - [Limits and allowed tools](/docs/limits.md) - [Authentication](/docs/authentication.md) - [Connect Claude Code](/connect/claude-code.md) ## Related guides - [One balance for your agent's tool calls](/ai-agent-tool-billing.md) - [Shared tools for a small team of agents](/ai-agent-tools-for-teams.md) - [Reduce the cost of your agent's tool work](/reduce-ai-agent-tool-costs.md) [Browse all agent tool guides](/agent-tool-guides.md) --- [HTML page](/ai-agent-tool-access-control) · [Agent index](/llms.txt)