Stack

One pick per harness component, and the accounts an agent acts through

Picks as of 7 Sep 2026. Re-derive their evidence from the repository withnode scripts/stack-evidence.mjs

Inside the harness

Harness

Needs the armory CLI · not on npm yet, build it from cli/ in the repository

Component	Pick	Score	Why	Install
Identity		99.92 signals	Behavior norms loaded from CLAUDE.md: think first, simplify, keep the diff smallTop of its shelf by score.	`armory install karpathy-coding-discipline --cli claude`
Memory	Contributed by Sentinel	99.53 signals	Stateful agents that edit their own long-term memory#10 on its shelf. Most rows above it are memory stores or graph builders outside the agent; in Letta the agent itself decides what to keep.	No one-command install · Source
Skills		99.93 signals	One bundle covering planning, review, testing and the rest of the lifecycleTop of its shelf by score.	No one-command install · Source
Tools		99.84 signals	Codegen, screenshot, PDF and trace commands for headless browser scripting#9 on its shelf. Picked to drive a browser from the terminal; most rows above it are terminal apps and other command-line tools, and Puppeteer, the closest match, has no WebKit.	No one-command install · Source
Hooks		99.12 signals	Intercepts ExitPlanMode so a plan is annotated and approved before work startsTop of its shelf by score.	No one-command install · Source
Sub-Agents	Contributed by Sentinel	99.72 signals	Plugin marketplace of sub-agents for Claude Code, Codex, Cursor and OpenCodeTop of its shelf by score.	No one-command install · Source
MCPs		99.93 signals	Issues, pull requests, repository search and commits without shelling out to git#3 on its shelf. Picked as the server most coding agents need first, published by GitHub itself; the rows above it drive a browser and convert documents.	`armory install github-mcp --cli claude`
Dispatch		99.22 signals	100-line graph framework; nodes and edges route work between agents#10 on its shelf. Picked as a library you drop into your own code: 100 lines with no dependencies. Above it, n8n is a platform you run, LangGraph, AutoGen, CrewAI, the OpenAI Agents SDK and Mastra are much larger frameworks, and ruflo, Orca and Symphony run coding agents for you.	No one-command install · Source
Evals		99.52 signals	Assertions and red-teaming for prompts and agents, runs in CITop of its shelf by score.	No one-command install · Source
Observability		99.63 signals	Traces, prompt management, datasets and evals in one self-hostable platform#3 on its shelf. The two rows above it are tracing add-ons to Sentry and MLflow; this is a whole platform on its own.	No one-command install · Source
Sandbox		99.84 signals	Firecracker microVM per run, ~150ms cold start, for untrusted code#2 on its shelf. The one row above it is Daytona, whose public repository stopped updating in June 2026.	No one-command install · Source

One command for these picks

11 picks · 2 in this command

armory install karpathy-coding-discipline --cli claude && \armory install github-mcp --cli claude

Not in the command, set them up from the source: letta-ai-letta, superpowers, playwright-cli, plannotator, wshobson-agents, pocketflow, promptfoo, langfuse, e2b-sandbox

Link to this stackhttps://armory-murex.vercel.app/stack

Accounts the agent acts through

These are accounts, not repositories, so they carry no Armory score; each line says how an agent gets access.

Slot	Pick	Deploy access
Computer	Orgo	ORGO_API_KEY, provisioned through bezalel.sh so the agent never sees the key
Sandbox	E2B	E2B_API_KEY, one microVM per run
Database	Supabase	URL and anon key to the agent; the service-role key is injected at the tool boundary, never into a prompt or a log
Email	AgentMail	API key and a per-inbox webhook secret (starts whsec_); verify the signature, or the route is an open relay
Phone / SMS / WhatsApp	Twilio	account SID, auth token and a purchased number; inbound requests are signed with X-Twilio-Signature
iMessage	Photon Spectrum	project id and secret, and the owner pairs their own handle with a one-time code (OTP)
Card / spend	Agent Card HQ	an identity check (KYC), then a funded balance; the slowest slot, so start it first
Payments / earn	Stripe agent-toolkit	a restricted key scoped to the agent, never the account key
Memory store	Supermemory	SUPERMEMORY_API_KEY, or self-host
Connectors	Composio	a key, plus the owner's OAuth consent per app; the agent never holds the third-party token
Deploy / hosting	Vercel	token and a linked project; after a CLI deploy, vercel alias set is a separate step, or nothing ships
Web / browser	Browserbase + Stagehand	key and project id, or local Chromium on the owner's profile, read-only

Start the slow slots first: the card needs an identity check (KYC) and iMessage needs the owner's one-time code, and both take days; then buy the phone number, then set up the email inbox and its webhook secret; everything that only needs a key takes a minute.

API returns both lists as JSON; Components lists up to three picks per component with its top-ranked rows.