# AI Consulting Workflows, Self-Hosted Software, and Context Sharing
- Session: 1609af18-9d5a-410a-9b9a-5e07ab097ed3
- Channel: Discord #đŁâcohort-voice
- Started: 2026-06-23T16:03:16.055Z
- Ended: 2026-06-23T16:39:28.974Z
- Participants: duckanbro, ECWireless, 0xHunter, louchi, Aphilos ⢠Pharo
- Tags: ai-consulting, codex-workflows, self-hosting, shared-memory, local-ai, mcp, frontend-qa, open-source-software
## TL;DR
The group interviewed 0xHunter about his AI-enabled consulting practice, self-hosted software strategy, Codex-heavy workflow, local model experiments, and the unresolved challenge of sharing scoped AI context across teams and clients.
## Summary
The meeting was a fireside-style interview in the RaidGuild cohort voice channel with 0xHunter, founder of Datalist Consulting. Hunter described building custom software, training, integrations, and technical consulting for medium-sized businesses such as medical, accounting, and other operationally complex clients. His companyâs thesis is that SaaS will become less dominant as custom software becomes cheaper to build, allowing smaller businesses to own their infrastructure, host their own tools, and eventually request or even make feature changes themselves.
Hunter explained that Datalist tries to self-host and internally build most of its own operational software, with exceptions such as email where reputation and deliverability make full self-hosting impractical. Internal examples included a CRM, transcription and summarization tooling, chat, and video calling. For new projects, the team decides case by case whether to build from scratch or start from open source. For simpler internal tools, Hunter may build directly with Codex. For more mature categories such as video calling or chat, he researches open source options and adapts them. He mentioned using La Suite, a French government open source suite, for video conferencing, and Mattermost for chat after testing multiple open source chat applications.
A major thread was how agentic coding has changed the work. Hunter said he has mostly switched from Cursor and Claude Code to Codex over the past few months, especially because Codex feels reliable for his workflow. He uses Codex with computer use capabilities for tasks that require browser or UI interaction, such as publishing to an app store, and described it as very effective for automating tedious web workflows. He also described using Codex and Cloudflare to help a nontechnical friend expose a 3D printer camera stream securely through a temporary Cloudflare tunnel.
The group discussed where AI workflows still feel risky or incomplete. Hunterâs biggest concern is QA, especially for frontend work. He can often trust Codex to write backend or functional code, but does not yet fully trust it to produce reliable polished frontend results without review. His ideal consulting workflow is to talk with a client, walk through their business processes, turn the meeting transcript into a software specification, make technical decisions with Codex afterward, and have the system reliably build the app. He sees this as plausible but not fully dependable yet.
Another key topic was context sharing. Duckanbro described RaidGuildâs interest in shared memory systems, task/workflow harnesses, and auditable organizational intelligence instead of isolated local agent sessions. Hunter agreed that context transfer between people remains unsolved. He sees teams and clients still copying and pasting Codex or Claude messages to share context. He framed the problem as one of scoping: deciding which conversations, projects, workspaces, and agent contexts should be accessible to whom at runtime. He speculated that MCP-style integrations could help, but only if authorization and scope are handled well.
The conversation also covered teaching clients to use AI tools. Hunter said some clients are now using Codex or Claude Code to make changes to internal tools and submit PRs that his team reviews. For interested clients, he teaches the underlying technical tradeoffs; for less technical clients, he prefers to build robust MCPs and controlled workflows so they can safely trigger actions without needing to understand all infrastructure details.
Toward the end, the group discussed local AI hardware and models. Hunter described using a business machine with 128GB of VRAM and personally running two RTX 3090s. He is especially interested in how model improvements can make existing hardware more valuable over time. He praised Qwen 3 62.7B as the first local model that met his bar for transcript summarization, and mentioned testing GLM 5.2. He also shared hunter.cafe, a local AI pet project that ingests PDFs or EPUBs, chunks books, generates wiki-style pages, checks for existing pages, and creates local images for those pages.
## Action Items
- Review Hunter's local AI project: Visit hunter.cafe to inspect Hunter's local AI wiki/book-ingestion experiment and consider whether any patterns are relevant to RaidGuild's memory or knowledge workflows.
- Evaluate open source video-call options: Explore whether self-hosted or open source video-call tools such as La Suite Meet could reduce Discord constraints for RaidGuild voice, transcription, and memory workflows.
- Deep dive on scoped context sharing: Use this interview as input for a research thread on organizational AI context transfer, including scoped memory, MCP authorization, shared agent workspaces, and auditability.
- Capture Codex computer-use workflow examples: Document practical examples from Hunter's workflow where Codex computer use handles browser or UI-heavy tasks, such as app store submission and Cloudflare tunnel setup.
## Notable Quotes
- 0xHunter: "Obviously SaaS is gonna die, software's gonna be cheap to build, and we wanna build that software for small companies and have them own it."
- Hunter believes cheaper AI-assisted software creation will let small businesses replace generic SaaS with owned, custom tools.
- 0xHunter: "My dream... is that I basically have a call with the client and we walk through the business... and then the transcript from my call would build something."
- Hunter's ideal consulting workflow is a client discovery call that becomes a reliable software build process.
- 0xHunter: "Something that hasn't been solved is context transferring context between people."
- Hunter sees shared context across agents, people, clients, and projects as one of the biggest unresolved problems.
- 0xHunter: "I think AI is gonna make agency size smaller... and it's gonna isolate people's workflows more and more."
- Hunter expects AI to shrink agency teams while increasing the need to reconnect otherwise isolated workflows.
- 0xHunter: "We've never seen... updating software has made hardware so much more valuable."
- Hunter noted that rapid model improvements can increase the practical value of existing local AI hardware.