Skip to content

Tools and agents register

What changed from the original design (Atul, 14 Oct)

Original design Now
Local models via Ollama (llama controller, deepseek/qwen coders) Deferred. Not used for now.
Llama 3.1 8B as controller/router Google GenAI Flash-Lite as the controller (Flash only for hard cases; see capacity register)
Local coding model Claude online, driven by the controller creating and updating GitHub issues; some development also on the GenAI API
Claude CLI as coder Not used in this phase; Claude (chat) codes directly in the repo and runs on the desktop
Small tool set Richer: GitLab runner, GitLab API, GitHub API, AWS CLI/API, collector, device link, Chrome
Other jobs on local models GenAI Flash-Lite (and Gemma for high-volume small jobs)

Agents

Agent Kind Used for Never does
Atul Person Approvals, merges to main, production, payments, sign-ins, DNS, UAT —
Ajay Person (legacy team) Production servers, crontab, DB facts; reached through the Google Doc only —
Claude (chat, MMDEP-PM) Claude online Programme manager; MMDEP architect and developer; reviews; release manager; judgment items Approve, merge to main, production, secrets
Claude online via GitHub issues Claude online Development tasks the controller files or updates as GitHub issues Merge to main
API/MM-Cron agent ChatGPT project, account A, hourly 07–23 IST API and MM-Cron Production without Atul
MWMC agent ChatGPT project, account A, hourly 07–23 IST Production website, MMCLONE sandbox Production without Atul
MCRM agent ChatGPT project, account A, hourly 07–23 IST MCRM, both streams Implementation before gates
WPDev agent ChatGPT project, account C WPDev website —
Ex-PM ChatGPT, attended only Governance repo housekeeping on request —
PE executor To be chosen Portfolio Engine development —
mmdep control GenAI Flash-Lite (Flash for hard cases) Triage findings into items, routing, summary and achievements log, meanwhile lines Approve, merge, production; never sees client data or secrets
GenAI helpers GenAI Flash-Lite / Gemma Classification, summaries, extraction, small development jobs Same limits as above

Tools on the laptop and online

Tool What MMDEP uses it for Access rule
Collector (_mmdep-tools\collector, to become mmdep collect) Hourly: API probes and NAV quality, CI/MR/issue state, SEO checks, weekly TLS certificates, question issues Read-only
GitLab runner (Docker, laptop,docker tags) CI for moneymarche and money-marche-testing groups, one job at a time Deploy jobs manual
GitLab API Read MRs, pipelines, issues Read-only for Claude (tokens act as Atul)
GitHub API (gh) Read and write MMDEP repo; read PE, MCRM Writes only to MMDEP docs/branches
AWS CLI / API Lightsail snapshots, later the tools server (DR-004) Read-only for agents; changes by Atul
Device link (Claude desktop app) Claude reads/writes connected folders and runs commands on the laptop Never touches mmdep_runtime or mmdep.db without an approved release
Claude in Chrome Reading GitLab, AI Studio, consoles as Atul Writes need per-action approval
MMDEP Actions panel (artifact) Atul's decisions, answers, summary Claude writes items and replies
GitHub Actions PE and MCRM CI on GitHub-hosted runners Minutes to be measured
Google Docs Ajay's question doc Edit in place