News Agent has no management forms — which means the management has to live somewhere else. It does: a set of commands with which the agent changes the settings, runs checks and prepares publications. This set is what makes the product manageable rather than “configured once and forgotten”.
Why an API
The API (programmatic access) is the language your AI agent speaks with the product. Through it, sources are added, thresholds set, rules and alerts defined, drafts prepared and checks run. A human does not need to learn this language: the agent speaks it for you, and you talk to the agent in ordinary words.
Knowing the API’s structure matters in two cases: when you work entirely without an AI agent, and when you connect the product to your own system.
The management key
The management key is set at installation in the NEWSAGENT_MANAGEMENT_KEY environment variable. What is stored is not the key itself but its irreversible fingerprint, so the original key cannot be recovered from the server’s settings.
- Until the key is set, changes are refused — any change request honestly answers that agent management is switched off. Reading keeps working.
- A request with the key gets the right to change everything — read, edit the settings, run checks and publications.
- The key is access to the whole product, so it deserves to be treated like an administrator password.
Who can do what
- A human in the interface — only look and mark: “read”, “favorite”. That is a deliberate restriction, not an unfinished feature.
- The agent with the management key — change the settings, run and pause the schedule, prepare drafts, repeat deliveries, clean the archive.
- Feed readers — receive what was published at the token-protected address, and nothing else.
The reference written for the agent
There is an address, /api/news/docs, that serves a description of itself: the list of commands, the description of every settings file and eleven ready agent playbooks — from “connect a source” to “assemble the morning digest”. The address is open without a key but shows nothing secret: it only explains how the management works.
The practical meaning: an agent that you gave the product’s address and key finds its way from there. No need to retell it the documentation — it reads the original.
The self-check
The address /api/news/health shows the product’s condition: whether the check schedule works, whether it watches the settings files, whether the delivery worker is alive, how many AI tasks stand in the queue and whether the database answers. It is the first address to look at on any “why the silence”.
What the agent has access to
- The schedule — the whole check plan in one request, a global pause and return to work, including an automatic return after a set time.
- An immediate run — a digest or a scheduled publication right now; a manual run does not shift the schedule.
- Re-processing — ask again for a summary or a materiality rating, complete an article’s full text.
- Work with material and events — merge repeats, hide an event from the feed, remove an item, pin a monitor’s baseline.
- Alerts — look at held messages, release them by hand, send a channel’s test message.
- Publications — assemble a draft, approve, send, unpublish, look at the version history.
- Settings — edits with validation, a backup and the journal; a return to the previous version.
- Maintenance — archive cleanup by retention, saving a page into the archive separately.
- Scripts — register your own script, run a test and read the run journal: what was called, how it ended, what it answered.
Rate and size limits
- No more than 120 changes per minute on a key — protection from an accidental storm of requests. On exceeding, the product answers “too often” and hints when to retry.
- Request size limits: settings — up to 256 kilobytes, material intake — up to 1 megabyte, a publishing brief — up to 16 kilobytes. A too-large request is rejected with an explanation.
- Reading is unlimited — as many views and health checks as you like.
Honest refusals
When something cannot be done, the product answers in words and explains the reason: “the draft was not created — no confirmed quotes were found in the material”, “the page was not found in the external archive”, “the channel is not verified — first send a test message”. There are no silences and no cryptic errors here: the answer always says what happened and what to do next.
One separate and important note: check not only that a request went through, but what the answer says. Honest refusals arrive as ordinary successful answers — with the explanation inside.
Live updates for the agent
The agent with the management key receives the same live notifications as the interface: an event appeared, material arrived, a delivery changed status, a publication is ready. That is how an external system can react immediately instead of polling the product every minute: “a new event — update the card in the CRM”.
Next → Chat & Commands