Autumn

Autumn is a research agent that runs in the cloud. You give it a brief, it searches, verifies and saves rows with sources, and you read the rows. A task takes minutes, not seconds, and a task is a conversation: you can answer its questions and refine it on the same task_id. Don’t use Autumn for a quick fact you can answer yourself.

Fastest path: the CLI

Install the Skill and sign in:

npm install -g @autumnai/cli
autumn setup
autumn login

Start a task and save its rows:

autumn run "Find recent funding rounds for climate software companies" --out rounds.csv

If the task needs an answer, ask the user and reply on the same task. If it is still working, resume waiting:

autumn reply <task-id> "Focus on North American companies"
autumn wait <task-id>

Exit codes: 0 completed, 1 failed or request error, 2 input required, 3 still working or approval pending, and 4 authentication required.

Install

npx skills add https://www.autumn.ai

This reads Autumn’s skill manifest and installs the skill wherever your agent keeps them.

Direct file

Point any agent that loads a skill from a URL at the file itself:

https://www.autumn.ai/skill.md

Claude Code project skill

mkdir -p .claude/skills/autumn
curl -L https://www.autumn.ai/skill.md -o .claude/skills/autumn/SKILL.md

Claude Code personal skill

mkdir -p ~/.claude/skills/autumn
curl -L https://www.autumn.ai/skill.md -o ~/.claude/skills/autumn/SKILL.md

Setup (once)

Base URL https://api.autumn.ai. Send X-API-Key: $AUTUMN_API_KEY (keys start with ak_).
No key yet? Follow https://www.autumn.ai/agent-quickstart.md: autumn login or two curl calls get a key once the user approves in their browser. Never ask the user to paste a key into chat, and never print, log or commit it.
Check it works: curl -sS https://api.autumn.ai/credits -H "X-API-Key: $AUTUMN_API_KEY".

The loop

# 1. Start
curl -sS -X POST https://api.autumn.ai/task -H "X-API-Key: $AUTUMN_API_KEY" \
  -H "Content-Type: application/json" -d @brief.json        # -> {"task_id": ..., "state": "submitted", ...}

# 2. Wait. Returns as soon as anything changes (max 50s). Repeat until "done": true.
curl -sS "https://api.autumn.ai/task/$ID?wait=50" -H "X-API-Key: $AUTUMN_API_KEY"

# 3. Answer or refine (queues if a turn is running, returns 202)
curl -sS -X POST https://api.autumn.ai/task/$ID/continue -H "X-API-Key: $AUTUMN_API_KEY" \
  -H "Content-Type: application/json" -d '{"message": "..."}'

# 4. Read rows (paged; use cursor for more). CSV keeps big results out of your context.
curl -sS "https://api.autumn.ai/task/$ID/output?format=csv" -H "X-API-Key: $AUTUMN_API_KEY" -o rows.csv

Every status response has: state, done, message (progress line, the agent’s question, or its final reply), progress (rows / target), credits_used, next.hint and app_url.

`state`	`done`	What you do
`submitted`	no	Wait. `queue_position` says where it is.
`working`	no	Wait. Relay `message` to the user as a one-line update.
`input-required`	yes	Show `message` (a question) to the user, send their answer with `/continue`.
`completed`	yes	Read `/output`. `outcome: partial` means it stopped short of the target.
`failed`	yes	Report `error.message`. `out_of_credits` is a hard stop: link https://platform.autumn.ai/?settings=usage.
`canceled`	yes	Stopped by `POST /task/$ID/stop`. Rows saved so far are still in `/output`.

Call the wait in a loop rather than sleeping. Each call is one update for the user.

How to research well with Autumn

Autumn plans from your brief. A vague brief doesn’t fail; it quietly answers a slightly different question. Write the brief like a contract:

Keep every specific the user gave you, verbatim. Names, URLs, geography, time window, exclusions, examples. Paraphrasing is where detail gets lost.
Say how many. "count": 25 (or a number in the prompt). Without one, “find companies” has no natural end.

Name the columns. If you know the shape, send a schema instead of describing it:

{"prompt": "Find 25 US fintechs that raised a Series A in 2025. Exclude crypto.",
 "count": 25,
 "output": {"kind": "research",
            "schema": {"company": {"type": "name"}, "domain": {"type": "domain"},
                       "round_date": {"type": "date"}, "lead_investor": {"type": "str"},
                       "source_url": {"type": "url"}}}}

State the per-row bar as rules. Rules are checks each row must pass, not goals: {"rules": {"series_a_2025": "Raised a Series A in 2025, with a source", "us": "Headquartered in the US"}}. Without a bar the agent decides what counts.
Pick the output kind. research for lists and tables. person or company for a deep profile per known entity. Pass the known entities as input_rows (e.g. [{"name": "Ada Lovelace", "company": "Analytical Engines"}]) or upload a file.
Pick the tier when it matters. "version": "ranger" for open-ended research where the agent must decide where to look or cross-check facts. "version": "scout" for bounded, well-specified work where speed matters. The tier is fixed for the life of the task.
Questions. clarify is false by default: the agent picks sensible defaults and runs. Set "clarify": true only when the user is present and an answer would change the work.

For complex or relayed requests, fetch GET /task/metaprompt. It returns the task spec schema and a metaprompt for turning a request into a spec for POST /task/start.

Iterate on the same task

More rows, a new column, a narrower list, fixing gaps: /continue the same task. Never start a new task for a follow-up; that loses the rows and the context.
Restate the goal as one clear instruction, not the user’s raw words. “also linkedin” -> “Add each founder’s LinkedIn URL to every existing row.”
After outcome: partial, look at what’s missing and ask for exactly that: “Find 8 more that meet the same rules” or “Fill lead_investor where it’s empty”.

Judge the result before reporting it

Report from /output rows, never from status text.
A shorter verified list beats a padded one. Autumn leaves a cell empty or unclear rather than guess. Say so, don’t fill it in yourself.
Spot-check a few rows against their source links for anything the user will act on.
Show the most useful rows, the count vs target, and the app_url for the full table.

Style

Terse and status-first. One line per update, e.g. working - 14/25 rows - verifying funding rounds.
Keep the task_id and reuse it.
Don’t show raw traces, internal errors or the API key.
Tasks spend credits. For a large or broad task, check GET /credits first and confirm with the user if the cost looks high.

Reference

API contract: https://api.autumn.ai/openapi.json
Docs index: https://www.autumn.ai/docs/llms.txt
App and API keys: https://platform.autumn.ai/?settings=connect
Help: founders@autumnai.com

Get started

Core concepts

Guides

Best practices

Use with agents

Skill

Autumn

Fastest path: the CLI

Install

Direct file

Claude Code project skill

Claude Code personal skill

Setup (once)

The loop

How to research well with Autumn

Iterate on the same task

Judge the result before reporting it

Style

Reference

​Autumn

​Fastest path: the CLI

​Install

​Direct file

​Claude Code project skill

​Claude Code personal skill

​Setup (once)

​The loop

​How to research well with Autumn

​Iterate on the same task

​Judge the result before reporting it

​Style

​Reference

Autumn

Fastest path: the CLI

Install

Direct file

Claude Code project skill

Claude Code personal skill

Setup (once)

The loop

How to research well with Autumn

Iterate on the same task

Judge the result before reporting it

Style

Reference