Fn2 Skill

Ask FN2 AI anything about public markets, then put the good questions on a schedule. From your terminal.

Get started · Run it yourself · In use · Commands · Scheduling · Security

Get started

claude plugin marketplace add fn2ai/fn2-skill
claude plugin install fn2@fn2

codex plugin marketplace add fn2ai/fn2-skill
codex plugin add fn2@fn2

FN2 AI runs autonomous research agents on public markets: earnings-call transcripts, prices and volume, SEC filings, FRED macro series, prediction markets, and the news around them. This skill hands that to your assistant, so Claude Code, OpenAI Codex, or Pi can launch an agent, read what it found, and put it on a cadence.

Things worth asking:

• Earnings calls. "What did NVIDIA say about data-center demand on the last call?" Answered with the executives' own words.
• Price action. "How is NVDA trading today, and what moved it?"
• Standing watches. "Notify me when AAPL drops more than 5% in a day."
• The daily brief. "Brief me on my watchlist every weekday before the open."
• Macro and filings. "Where is the yield curve now versus a year ago?" · "What changed in TSLA's latest 10-Q?"
• What the market is pricing. "What odds are prediction markets giving a Fed cut this year?"

It is one dependency-free Python 3 file tree: no pip install, no build step, standard library only.

Install details

The one-command install is under Get started. Pi, or local development in any harness: clone the repository and link the canonical skill directory.

git clone https://github.com/fn2ai/fn2-skill ~/.fn2-skill
ln -s ~/.fn2-skill/skills/fn2 "${HERMES_HOME:-$HOME/.hermes}/skills/fn2"

make install from a checkout links the canonical skill into Claude, Codex, and Pi at once.

A plugin installs a skill, not a binary, so nothing lands on your PATH. Usually that is fine: you talk to your assistant, not to a shell command. If you want fn2 in your own terminal too, run this once (or just ask your assistant to "put fn2 on my PATH"):

python3 "${CLAUDE_CONFIG_DIR:-$HOME/.claude}"/plugins/cache/fn2/fn2/*/skills/fn2/scripts/fn2 link

That writes a small launcher into the first writable bin directory on your PATH and tells you if there isn't one. The launcher re-resolves the skill on every run, so plugin updates do not strand it. fn2 unlink removes it.

Once linked, fn2 auth login is the same browser device flow as asking your assistant to connect: code in the terminal, approval at fn2.ai/start, credential in the Keychain.

Running it yourself

Every screenshot in this README is a command you can run once fn2 link has put the launcher on your PATH. A few things make it quick to live in:

fn2 agents                       # ids print 8 characters wide...
fn2 status ebab1623              # ...and 8 characters is enough to refer back
fn2 status "market open"         # so is a name
fn2 result ebab1623              # newest run, no run id to copy
fn2 start "..." --wait           # ask and get the answer in one command
fn2 run ebab1623 --wait          # trigger and block until it finishes

Ambiguous references are refused with the candidates listed, never guessed. --wait uses the API's own long-poll rather than hammering it, and takes --timeout (default 300s); Ctrl-C is always safe.

Output is colourised when it is going to a terminal and plain everywhere else. --color never and NO_COLOR both turn it off; --json is never colourised. On a console that cannot encode ✓, the glyphs fall back to ASCII instead of raising UnicodeEncodeError.

In use

Standing watches, in one sentence.

fn2 agents, the whole fleet

Every agent, its cadence, and when it last ran.

fn2 status <id>, one agent in full

Cadence, model, the prompt it runs, and its recent runs.

fn2 result <id> <run>, read the research

The finished briefing, in the terminal, with token accounting.

fn2 start --when, cadences in words

"market close", "weekdays at 7am", "before market open".

fn2 whoami and fn2 models

Plan, quota, device scopes, and the models you can run on.

fn2 permissions and fn2 doctor

What the credential cannot do, and a one-command health check.

Every screenshot is generated by running the real CLI. See tools/screenshots.

Commands

Any command that takes an agent takes an id, an id prefix, or a name.

Every command accepts --json for scripting. Run results are available via fn2 result and in the FN2 AI web app; add --email at creation to have each completed run delivered to your verified address.

Scheduling in everyday words

--when accepts ordinary phrasing:

before market open      every weekday 09:00 ET (30 min before the bell)
market close            every weekday 16:05 ET
weekdays at 8:30am      cron under the hood, your timezone
daily / hourly          named cadences
before next market open a single run at the next open
in 2 hours / tomorrow at 9am / 2026-07-22T13:00    one-off runs
cron:0 9 * * 1-5        raw cron for full control

Market phrases default to America/New_York and resolve against the exchange calendar server side, so they skip holidays and respect half-days. See skills/fn2/references/scheduling.md.

Security and configuration

The credential is resolved in this order, first hit wins:

1. FN2_API_TOKEN environment variable (canonical headless override; accepts fn2d_… device tokens and fn2_… API keys)
2. FN2_API_KEY environment variable (legacy alias)
3. macOS Keychain (service fn2-skill; used automatically on macOS)
4. ~/.config/fn2/credentials.json (written 0600, directory 0700)

There is no .env support, so a working-directory file can never inject a credential. Other knobs: FN2_BASE_URL (default https://fn2.ai), FN2_CONFIG_HOME.

Hard safety rails:

• Origin pinning. The API base must be https:// on fn2.ai or *.fn2.ai; the credential is never sent anywhere else (FN2_ALLOW_DEV_BASE=1 exists for local development only and warns loudly).
• Redaction. The credential appears only as a masked fn2d_…wxyz form in all output, including --json and error text.
• Revocation. Revoke any connected device in the FN2 AI web app under Preferences → Developer → Connected Devices (or fn2 auth logout --revoke); it takes effect within a few minutes everywhere.

Found a security issue? See SECURITY.md.

How it works

skills/fn2/scripts/fn2            → entry point (adds scripts/ to sys.path, calls the CLI)
skills/fn2/scripts/fn2cli/
  config.py            → credential + endpoint resolution, origin pinning, redaction
  storage.py           → macOS Keychain / owner-only file + pending-login state
  deviceauth.py        → the fn2.ai/start device-authorization flow
  client.py            → urllib HTTP layer + typed errors
  api.py               → one function per FN2 AI endpoint
  schedule.py          → "before market open" into an FN2 AI schedule object
  render.py            → human / JSON output
  style.py             → ANSI colour + glyph fallback, off unless stdout is a terminal
  cli.py               → argument parsing + dispatch

It talks only to the documented public API (/api/v1/agents, /runs, /usage, /models, /skill/device-auth/*) with your scoped credential. No admin access, no scraping. See skills/fn2/references/api.md.

Development

make test        # unittest, no network
make lint        # byte-compile every module
make validate    # strict Claude plugin validation (when Claude Code is installed)
make install     # link the canonical skill into Claude, Codex, and Pi
make media       # regenerate docs/media (needs Chromium + ImageMagick)

Contributions welcome. Start with CONTRIBUTING.md.

License

MIT, see LICENSE. Bundled documentation fonts (JetBrains Mono, Libre Franklin) are SIL OFL 1.1; see tools/screenshots/fonts.