Driving Archon with the Companion & Terminal

0The mental model

Two things in this app are both called by familiar names. Untangle them once and the rest is easy.

A column is a step, and its type decides what runs when you hit Run:

inputai tooltransform cloud

• input — data you typed or imported (company name, domain).
• ai — runs an LLM prompt per row (Claude / OpenAI / OpenRouter).
• tool — calls an enrichment API per row (find email, scrape site, Apollo…).
• transform — cheap local cleanup (first name, domain-from-email).
• cloud — offloads heavy work to the Trigger.dev cloud.

Every cell (row × column) goes pending → running → complete and accrues a tiny cost. The dollar figure you see is the running total.

1The window, mapped

Three regions. Left = what you have. Center = what you're looking at. Right = how you drive.

2The one golden rule

AThe Terminal — how we drive Archon

The Terminal tab is not a bash prompt. It launches the Claude Code CLI (claude) directly, running inside the gtm-orchestrator checkout and wired to the table you have open. You talk to it in plain English; it does the work. This is the default way to use Archon.

What's wired up for you

• Your selected table — the active table id is passed in (ARCHON_WORKFLOW), so archon context already targets it. Switch tables in the UI → the Terminal re-attaches.
• The same database the app uses (ARCHON_DB) — anything the agent writes shows up live in the center grid, and vice-versa.
• The archon CLI on PATH — the agent's tool for reading and writing the table, running enrichment, and clearing gates. Run archon help to see every verb.
• The orchestrator playbook — the checkout's own CLAUDE.md, skills, and slash-commands are loaded, so the full GTM workflow is available, not just table edits.

The enrichment flow

However you phrase it, a run walks the same path. Either start from a recipe (columns pre-built) or have Claude design them with the plan flow, then run and clear any gate:

1. archon context — read the table: name, columns, row count, a sample. Always first.
2. archon table create … --recipe company-enrichment or archon plan context → plan propose --json → plan apply — get the enrichment columns in place (the plan flow is the same engine as the old "Build columns" button).
3. archon run --scope all --now — execute every cell. --scope sample tries a few first; estimate prices it before you commit.
4. archon gate approve <file> --signature YOU — clear any checkpoint the engine raised.

The full archon surface

From archon help — the real verb list:

archon context                                  # table, columns, row count, sample rows
archon clients | client create --name N
archon tables | table create --client <id> --name N [--recipe <id>]
archon recipe list                              # company-enrichment, icp-scoring, competitor-scan, hiring-signal…
archon rows list [--limit N] | rows read <rowId> | rows add --data '{"col":"v"}'
archon cell write <rowId> <colId> <value>
archon column add --name N --type input|function|tool [--tool id --params '{"p":"colId"}']
archon plan context | plan propose --json '<spec>' | plan apply <draftId>   # design ai/tool/waterfall columns
archon run [--scope all|pending|sample] [--n N] [--now]      # the Run button
archon estimate [--scope …]                          # pre-run cost in cents
archon expand --source-col <id> --target <wf> --field-map '{item:col}'   # JSON array → one row each
archon tool list | tool add <file.json> | tool test <id> --params '{…}'
archon gate list | gate show <file> | gate approve/reject <file> --signature YOU

BEnd-to-end: enrich inherited-mess

The table has 3 rows and one Company column. Goal: add a domain, a one-line description, and a decision-maker email, then run it. You select the table in the UI, open the Terminal, and say what you want — here's the conversation that follows:

> enrich this table: add each company's domain, a one-line description,
  and a decision-maker email — then run all 3 rows

Reading the table first.
⏺ archon context
✓ "inherited-mess" · 3 rows · columns: Company

I'll design the columns with the plan flow, then apply and run.
⏺ archon plan context                       # entity fields + the real tool ids
⏺ archon plan propose --json '{ "columns": [ … ] }'
✓ draft d1a2 — Domain (tool), Description (ai), Email (waterfall) · est $0.07
⏺ archon plan apply d1a2                     # instantiates the columns (expands the waterfall)

Columns are in. Running every row now.
⏺ archon run --scope all --now
✓ 3 rows · 9 cells · complete · $0.07

Done. The engine raised a checkpoint for the contact stage — approve it?
> yes, sign it Mitchell
⏺ archon gate approve .checkpoints/04d-enrich-contact.json --signature Mitchell
✓ gate cleared

You typed two English sentences; Claude Code did the rest. The grid in the center fills in live as each cell completes.

CThe Companion — the old panel legacy

The Companion is the same agent in a graphical panel instead of the Terminal. It still works, but we're standardizing on the Terminal — so reach for it only if you specifically want the GUI affordances below.

Everything here is reachable from the Terminal by just asking — which is why it's the default now.

DWhich surface to use

ECheat sheet