Watchtower

Watchtower

Tech job monitoring for AI agents. Say what you are looking for once and get only the new postings that match, from the job boards of tech companies and startups, as structured JSON.

Agents are bad at waiting for a job to be posted. A session lasts minutes, careers pages are heavy to re-read, and "has anyone posted an iOS role in Austin yet?" turns into re-running the same search every day. With Watchtower the agent creates a watch once, in plain language:

{ "query": "iOS jobs in Austin making at least 150k a year with a maximum of 6 years of experience" }

Watchtower checks the job boards of tech companies and startups on a schedule and remembers which jobs were open. get_changes then returns only the new postings that match, and a webhook can wake the agent when one arrives. An agent can also watch one company's board by URL and get JOB_ADDED, JOB_REMOVED and JOB_UPDATED events.

• No URL needed. A watch created from a query covers every board Watchtower monitors: a built-in directory of tech company and startup boards plus every board anyone has watched by URL.
• Pay and experience. Jobs carry salary and experience_years when the posting states them, and watches filter on min_salary and max_experience_years.
• Greenhouse, Lever, Ashby, Workable, SmartRecruiters, Recruitee, Workday, iCIMS, Oracle Recruiting, Eightfold and SuccessFactors are read through each platform's own endpoints: no bot walls. So are Apple (jobs.apple.com), Google (careers.google.com), Amazon (amazon.jobs) and Microsoft (careers.microsoft.com), which run their own careers sites. Any other careers page works if it publishes schema.org JobPosting markup.
• Filters on the watch, so get_changes only returns what matters: keywords, all_keywords, exclude_keywords, locations, seniority, remote_only, min_salary and max_experience_years. Every job carries derived remote and seniority fields.
• Many companies in one call: pass urls instead of url.
• MCP (Streamable HTTP) and REST, backed by the same service layer.
• Free and anonymous: a client gets a token and up to 50 watches (a search across all boards is one watch).
• TypeScript, Node.js 22, Fastify 5, PostgreSQL, the official MCP TypeScript SDK. No LLM or third-party API keys required.

Use the hosted service

Watchtower runs at [watchtower.lat](https://watchtower.lat), free, with no sign-up. MCP endpoint: https://watchtower.lat/mcp.

• Claude Code: claude mcp add --transport http watchtower https://watchtower.lat/mcp, or install the plugin, which adds a skill that tells Claude when to use it:

  /plugin marketplace add connorlagana/watchtower

/plugin install watchtower@watchtower

- **Claude.ai / Claude Desktop**: Settings → Connectors → Add custom connector → `https://watchtower.lat/mcp`.

• Cursor, VS Code: one-click buttons at watchtower.lat/#install.
• Anything else: { "mcpServers": { "watchtower": { "type": "http", "url": "https://watchtower.lat/mcp" } } }.

Listed in the MCP Registry as lat.watchtower/watchtower (server.json).

Quick start

cp .env.example .env

docker compose up --build # app on http://localhost:3000, Postgres alongside docker compose --profile demo run --rm demo # the end-to-end demo below

Without Docker (Node 22 and a running Postgres):

npm install

export DATABASE_URL=postgres://postgres:postgres@localhost:5432/watchtower npm run migrate npm run dev # or: npm run build && npm start

Demo

npm run demo (with DATABASE_URL set) runs the core loop end to end against a local fixture careers page:

1. create client → 2. create a job watch with keyword ios (takes the initial snapshot) → 3. a second agent watches the same board and reuses the same resource → 4. a third agent names no board and creates a search watch from "remote iOS jobs paying at least 150k with a maximum of 6 years of experience" → 5. a re-check where only the page around the jobs changed (session token, "N minutes ago") reports no change → 6. the source adds an iOS job and an Android job → 7. Watchtower checks and detects the change → 8. an MCP client calls get_changes:

{

"changes": [ { "type": "JOB_ADDED", "summary": "New job: Senior iOS Engineer (Remote - US)", "data": { "job": { "title": "Senior iOS Engineer", "location": "Remote - US", "company": "Acme Robotics", "url": "…/careers/ios-303" } } } ], "cursor": 2, "has_more": false }

The Android job is filtered out by the keyword. 9. A second get_changes call returns []. 10. The second agent, with no keyword, sees both new roles. 11. The third agent gets the iOS role with the salary and experience read from the posting.

The demo runs its fixture on 127.0.0.1, so it turns on ALLOW_PRIVATE_NETWORKS for its own process only.

Connecting an agent (MCP)

{ "mcpServers": { "watchtower": { "type": "http", "url": "http://localhost:3000/mcp" } } }

watch_jobs also accepts webhook_url for push delivery (see below).

The tool descriptions and server instructions tell agents to prefer Watchtower over re-running job searches or re-checking careers pages.

Search watches

A watch with no url is a search watch. It is its filters, and it reads the new postings of every monitored board.

curl -s -X POST localhost:3000/v1/watches -H "Authorization: Bearer $TOKEN" -H "content-type: application/json" \

-d '{"query":"iOS jobs in Austin making at least 150k a year with a maximum of 6 years of experience"}'

{

"scope": "all_boards", "interpreted": { "filters": { "keywords": ["ios"], "locations": ["austin"], "min_salary": 150000, "max_experience_years": 6, "include_unknown": true }, "notes": [] }, "coverage": { "boards": 1705 }, "matching_jobs_count": 3, "current_jobs": [ { "title": "Senior iOS Engineer", "company": "Acme", "location": "Austin, TX", "salary": { "min": 165000, "max": 210000, "currency": "USD", "period": "year", "annual_min": 165000, "annual_max": 210000 }, "experience_years": 5, "url": "…" } ] }

• The query is read by rules, not a model. Watchtower still needs no LLM or API key. The reading comes back as interpreted, with notes for anything it could not use, so the calling agent can check it. Explicit filters always win over the query.
• Role words become keywords (any of, for "iOS or Android") or all_keywords (all of, for "data scientist"). Generic words such as "engineer" and "developer" are dropped when a more specific word is present.
• "in Austin", "in Austin, TX", "in New York or remote" become locations. A bare "remote" becomes remote_only.
• "at least 150k", "$180,000+", "$45/hr" become min_salary, converted to a yearly figure.
• "a maximum of 6 years of experience", "3-5 years", "I have 4 years of experience" become max_experience_years.
• "senior", "staff", "entry level" and the other levels become seniority. "no managers" becomes exclude_keywords.
• Coverage is tech companies and startups, not the whole internet. The built-in directory (src/search/boards.ts) lists 1,705 company boards on Greenhouse, Lever, Ashby, Workable, SmartRecruiters, Workday, iCIMS, Oracle Recruiting, Eightfold and SuccessFactors, plus the careers sites of Apple, Google, Amazon and Microsoft, from seed-stage startups to large employers in every industry that hire developers. Each was confirmed against its platform's API with open jobs. Every role those companies post is covered, not only engineering. The live list is published at /companies and through list_companies. On top of that, every board any client watches by URL is covered, and stays covered (see Growing the directory).
• Pay and experience come from the posting. Lever, Ashby, Greenhouse, Recruitee and JSON-LD pay fields are read directly; otherwise the posting text is parsed ("$150,000 - $200,000/yr", "5+ years of experience"). A job passes min_salary when the top of its range reaches it, and max_experience_years when it asks for no more than that.
• Postings that state neither are still reported, without a salary or experience_years field, because many postings state no pay. Pass include_unknown: false to report only postings that state a qualifying value.
• Only new postings are reported (JOB_ADDED), and only those that appear after the watch was created. current_jobs on creation and on get_watch is the baseline of what is open now.
• SmartRecruiters, Workday, iCIMS, Eightfold (except older-endpoint tenants), SuccessFactors (except RSS sitemaps), Google and Microsoft listings carry no posting text, so their jobs never have salary or experience_years. Apple and Oracle Recruiting listings carry only a short summary, which rarely states either.
• Google jobs have no location and only the first part of the title (see Google), so a search with locations never matches them.

Growing the directory

The directory grows in three ways.

• Agents grow it by using it. When a client watches a platform board by URL and it has open jobs, the board joins the directory and stays there after that watch ends, so every search watch covers it from then on. At most INDEX_MAX_PROMOTED (5,000) boards are kept this way, and one that fails eight checks in a row is dropped. Arbitrary careers pages are never kept; only boards on the platform APIs.
• A careers page works as a way in. watch_jobs with https://acme.com/careers usually finds no job markup there, because the page only links to or embeds the company's board. Watchtower the