Coding agent control plane

Invoker

Know what your coding agents are doing before it ships to prod.
Watch, pause, retry, approve. All from Desktop or Slack.

Get early access

Star Invoker on GitHub

See every runApprove before mergeReplay the trailRetry safely

Invoker desktop showing a scrubbed coding-agent workflow with readable task nodes, status rails, and dependency edges

Product demo

See agent work before you ship it.

Track every run, branch, attempt, approval, and retry from one control surface. When work needs judgment, step in from the desktop or Slack.

See every coding-agent run, selected attempt, and branch lineage in one control plane from desktop or Slack.

Trusted by engineers from

Product

Agent work you can inspect before it ships.

inspect, redirect, and approve agent work

Navigate a real DAG

Open any completed task and see the AI session, branch, and evidence that produced it.

Architecture

The work is durable. The choices stay human.

Agents can make mistakes. Don't let that hurt you. Know what happened.

Invoker is a workflow system, not a chat wrapper. You are in control, not AI. Our architecture does not permit AI to run wild. Save time, tokens, and money.

Plan

A plan captures the tasks, dependencies, prompts, commands, and review rules before work starts.

Control

Invoker writes state before action, checks what is ready, and resumes through the same path after restarts.

Execute

Each attempt runs in isolation and reports its branch, logs, evidence, and outcome.

Review

Human gates, merge checks, retries, and downstream updates keep shipped changes intentional.

workflow events

00:00workflow.createdrelease-plan.yaml parsed

00:04task.readyapi-contract and test-fix queued

00:17attempt.startedisolated worktree allocated

01:26attempt.completedbranch and summary persisted

01:29review.gateapproval required before merge

01:44downstream.readydocs and changelog unblocked

Operations

When agents fail, you still know what happened.

The hard problem is not launching more agents. It is seeing what they did, recovering from partial failure, retrying through the same control path, and proving review happened before merge.

Persisted recovery modelState is written before action, so recovery can re-check durable state and resume through normal command routes.

Selected attempts and lineageAttempts, branches, logs, and selected outputs stay inspectable so humans decide what moves forward.

Executable review evidenceReviewers get the branch, attempt summary, command logs, readiness reason, and verification result.

Slack-native controlStart plans, check status, approve, reject, and retry from Slack without leaving the conversation.

Security and governance

Keep agents inside the lines.

Invoker does not make agents responsible. It makes their work controllable: isolated attempts, one mutation path, explicit gates, and evidence reviewers can inspect.

product principlespublic-safe