Self-Hosted Workflow Automation

As of June 2026, AIDiveForge tracks 9 self-hosted workflow automation. Curated self-hosted workflow automation tracked by AIDiveForge. Listings are verified against each tool's live website and re-checked regularly.

Last updated June 12, 2026 · 9 tools

1. 

   1. AI Mime

   AI Mime records a macOS task once, then compiles the raw trace into a coordinate-free skill: deterministic scripts where possible, a browser harness or native UI agent only at decision points where necessary. The self-healing loop is the real differentiator — when a run fails, an agent reads the logs, triages the issue, and patches the skill instead of silently dying. The output is a readable directory of files, not a locked binary, so Claude Code or Codex can call it directly. The wall appears on Windows and Linux: this is macOS-only, and teams needing cross-platform coverage will hit that ceiling before the third workflow.

   FreeOpen Source
2. 

   2. Freu AI

   Freu AI's approach is observe-once, compile, execute-forever: a human performs a workflow, the agent records and compiles it into a locally-runnable program, and from that point forward execution runs without calling a model on every step. The vendor positions this as the core cost argument — token spend happens during the learning phase, not during the thousands of subsequent runs. That architecture fits invoice routing through ERPs, clinical evidence extraction, and batch record migration across legacy systems that have no API surface. The wall appears when a workflow changes: any meaningful UI or process shift requires a new learning pass, which means ongoing human expert time isn't eliminated, just front-loaded.

   Paid
3. 

   3. GhostUser

   Each persona — a cautious newcomer, a skeptical evaluator, a power user, a time-pressured visitor, a motivated buyer — navigates your app autonomously, flags where it gave up, and logs why. Console errors, failed network requests, and 5xx responses get caught in the same pass, so you get UX feedback and QA signal in one run. It connects directly to localhost, which means you catch issues before they leave your machine. The tool runs on your Claude API key, so cost scales with usage rather than with a seat count. Where it breaks: the feedback reflects what five hardcoded personas notice, not the distribution of your actual users.

   FreeOpen Source
4. 

   4. Gumloop

   Gumloop lets growth, sales, and ops teams wire together multi-step AI agents that run on their own — pulling from external APIs, enriching CRM records, drafting content, and firing results into Slack or Teams without a human trigger per run. The visual builder handles the common cases well: lead enrichment, meeting prep, competitive research. Branching logic that depends on what a previous step returned is where the ceiling appears — complex conditional paths push teams toward adding custom code nodes, which means they are now maintaining two layers. Security and compliance teams get enterprise-grade controls over AI usage, which matters when rolling out to non-technical employees at scale.

   Paid
5. 

   5. Onpilot

   The platform connects agents to ERP, CRM, support tools, and custom APIs, then layers in approval steps, permission scopes, and audit logs so the agent cannot act unilaterally on sensitive operations. Agents can search, reason, take action, and hand off to a human — the approval step pauses execution and sends an interactive Slack message before anything ships. Multi-tenant architecture means a single deployment can serve isolated customer or plant workspaces with per-tenant access control. Where it breaks: Onpilot is a custom-built, consultative engagement, not a self-serve platform you configure over a weekend — teams without clear workflow documentation will stall during scoping.

   Paid
6. 

   6. Ornold MCP

   The structured data describes a browser automation platform for parallel antidetect workflows, vision-first interaction, and CAPTCHA solving at scale. However, the scraped page content is from an unrelated travel-identification app called Spotter. There is no factual basis from the page to describe how the tool handles parallel execution, how its AI agent layer interprets natural-language task definitions, where its CAPTCHA solving hits rate limits, or when the free tier stops being sufficient. Publishing claims without a sourced page would mean fabricating production details — the one thing an engineering lead or PM cannot afford to act on.

   Paid
7. 

   7. RiddleRun

   RiddleRun combines a CLI and an optional self-hosted web app, both running inside Docker, so your test environment travels with the repo rather than living on someone's laptop. You define a user journey in JSON — steps, assertions, expected outcomes — and a Playwright/browser-use agent executes the whole sequence autonomously. The Docker-first setup means teams can wire it into CI without installing a browser stack on the build machine. The project has two GitHub stars and one open issue at the time of curation, which signals early-stage maturity — documentation depth and community support are thin, and the agent's decision logic is largely a black box to the teams running it.

   FreeOpen Source
8. 

   8. SoMatic

   The core workflow is a CLI command that takes a screenshot, runs element detection locally, and returns numbered marks with coordinates as JSON — so agents target elements by ID, not by fragile pixel hunts. Every action returns JSON, which means downstream agents can chain steps without parsing unstructured output. The self-hosted, MIT-licensed model runs on your own hardware, so no screenshot data leaves the machine. The wall appears with non-standard or highly dynamic UIs where YOLO detection misses elements or mislabels them — teams handling those cases add a fallback coordinate layer manually. At this GitHub star count, the community size is small, which means debugging edge cases happens in the codebase, not a forum.

   FreeOpen Source
9. 

   9. Yansu

   Yansu, from Isoform, flips that contract: it watches how work actually gets done, learns the pattern, and builds the automation from observation rather than instruction. The vendor describes autonomous loop-based execution across desktop tasks, support ticket handling, and form-filling — with a local-first processing model that keeps data off third-party servers. Teams capturing tribal knowledge get the most direct value here; the agent surfaces patterns that live in no documentation. The ceiling appears when workflows require branching logic or cross-system integrations that go beyond what observation can infer, at which point teams are back to configuring manually. No public API is available, which limits how far this plugs into existing engineering stacks.

   Paid

Listings on this page are sourced and verified by the AIDiveForge data pipeline. AIDiveForge is editorially independent — no money changes hands for inclusion.