introduction

Documentation Index

Fetch the complete documentation index at: https://docs.invokehq.run/llms.txt

Use this file to discover all available pages before exploring further.

Invoke sits between your agent and the real tools it calls — Stripe, CRM systems, Slack, Linear, MCP servers, internal APIs. When your agent decides to take action, Invoke controls what happens next: validating scope, classifying the risk, reconciling ambiguous outcomes, and returning a trace of every decision.

​

Why agent execution fails in production

AI agents are good at reasoning. Production execution is a different problem. Three failure modes cause real operational damage after the model makes a correct decision: API timeouts with unknown outcomes. Your agent fires stripe.charge_customer and the connection drops. The charge may have succeeded. Retrying blindly can move money twice. The model cannot reason its way out of this — it needs runtime-level outcome reconciliation. State drift mid-run. The customer record, deployment status, or invoice the agent resolved at plan time is no longer current truth when the write executes. An agent that proceeds anyway touches the wrong entity. Duplicate retries. A lost response from a write-once API — an issue creation, a refund, a CRM update — results in a second side effect landing when the first response is replayed or retried without idempotency enforcement.

​

What Invoke does

Invoke resolves each failure mode before a side effect reaches production:

• Classifies every tool as idempotent, write-once-checkable, irreversible, or approval-bound so execution behavior is automatic
• Writes a commit record before touching the outside world, giving every action a durable execution ID
• Reconciles unknown outcomes by checking live source-of-truth state before allowing a retry
• Gates risky actions under policy, scope, and approval rules tied to a stable agent identity
• Returns a trace with policy decisions, attempts, idempotency state, and final outcome

Your agent calls Invoke instead of every API directly. The team gets structured proof of what happened — not logs, a real execution record.

​

Works alongside your existing stack

Invoke does not replace your agent framework. Keep your LangChain chains, Claude Code tools, OpenAI function calls, MCP servers, and custom scripts. Invoke sits in the execution path when the agent is about to touch the world.

​

Where to go next