Workix

Workix MCP

Runtime source catalog: the current platform list is platforms.json. The older platform research matrices are historical evidence, not an active integration list. include_jobs is active; include_agent_gigs is active for structured agent task sources; discovery_only records external agent APIs/MCPs for networking and contact discovery without pretending they are job feeds; watch_pending is explicit-test only; watch_browser is manual browser-only.

Local MCP server for AI agents (Cursor, Claude, etc.) that work with the Workix hub and freelance boards.

What's new in 1.2.6

• First public release since 1.2.2 — bundles session-flow guidance (1.2.3), the RU browser-profile adapters Profi.ru (услуги) and Avito Работа (вакансии) (1.2.4), and X (Twitter) hiring/founder leads (1.2.5). Browser-profile sources read a logged-in profile and never apply — отклик остаётся ручным; Avito and X are double-gated (AVITO_ENABLE=1 / X_ENABLE=1) and opt-in only.
• Agent-community catalog. The Colony, Agent Community, Moltbook, Chirper.ai, SocialAIA, Get Posting Board and 0xWork are listed in platforms.json for networking and contact discovery.

What's new in 1.2.2

• Participants and mentors. Public hub copy now describes people as participants: specialists, collaborators, founders, and mentors can connect through skills and interests. The API tool names and legacy performer URLs remain unchanged for compatibility.
• Collaboration matching. Participant cards can record networking, startup, open-source, and equity preferences under collab; workix_list_performers can filter by those preferences.
• Profile visibility and bump. hidden controls whether a participant appears in feeds/search, while workix_bump_profile resurfaces a card no more than once every 3 days.

What's new in 1.2.0

• Collect → search split. workix_collect ingests every source into the store in parallel (HTTP/RSS/boards/HH + a full Telegram channel sweep); workix_db_search then ranks/filters the store across all platforms with no network — word-boundary keyword match, résumé/"for hire" drop, cross-post collapse, per-advertiser cap, and an "already applied" cross-check against the outreach log.
• Telegram two modes. workix_tg_search now takes mode: search | dump — server-side per-term walk vs. sweep-then-match-locally. Fixed the since window (Telegram-only checkpoint, 2-day floor).
• New sources: web3.career, reactjobs.io (React/RN/Flutter/mobile), and aijobs.net (AI/ML) adapters; ATS company list expanded +17 (Proton, Tailscale, Sierra, Cognition, Decagon, ElevenLabs, Perplexity, RevenueCat, Zed, Modal, Baseten, …); YC Work at a Startup and Wellfound read through a logged-in browser profile (scripts/board-open.mjs + board-save.mjs).
• Telegram channel lists (telegram-channels.json / .example.json) synced to 80 channels, dead entries pruned.

Previously (1.1.0): adapters for JobSearchDB, regional boards, Remocate, Startup Jobs MCP, Startupium, TheHub; Telegram checkpoints / incremental since / scan timing / source-quality rotation; runtime source catalog and access docs.

Full release notes: CHANGELOG.md.

Workix is a shared catalog where projects (startups/products/communities), roles & orders, participants, and mentors find each other. Search by skills and interests, then connect through the owner’s preferred contact or apply flow. It is not a closed freelance marketplace: agents use this MCP to search that catalog, manage your listings, and also pull opportunities from external boards while keeping platform credentials on the user’s machine.

Install (recommended): clone this repo → mcp/ → npm install && npm run build · Docs: workix.co/agent · api.txt · llms.txt Optional shortcut: npx -y @workix/mcp (npm may lag behind git).

Where it is published

Dataset

workix/workix-mcp — open dataset generated from this server's own tool definitions (Apache-2.0), in three configs:

from datasets import load_dataset

qa = load_dataset("workix/workix-mcp", "qa", split="train")

Regenerate after changing the tool surface — it reads tools/list off the running server:

node scripts/dump-tools.mjs

HF_REPO=workix/workix-mcp node scripts/dataset-build.mjs

Sources: scripts/dump-tools.mjs · scripts/dataset-build.mjs · seeds in scripts/dataset-qa.json

What you can use it for

• Discover — search hub projects, roles, orders, participants, and mentors relevant to skills or interests
• Publish / update — create or edit a startup or role, maintain a participant profile (WORKIX_AGENT_KEY)
• Freelance workflow — digests and search across supported boards, draft proposals, prepare browser apply steps (submit only with human confirmation)
• Ship a product card — publish a live site/PWA URL into the dStore catalog and find similar apps

• Hub tools — search/list/create projects, roles, profile; feedback/support (needs WORKIX_AGENT_KEY)
• Board tools — digest / search / draft / prepare-apply on Upwork, Freelancehunt, Kwork, FL RSS, … (credentials stay in your env)
• Downloadable adapters — heavy board modules are not required at install; MCP downloads them from the hub registry on first use and caches locally

Open contribution: see CONTRIBUTING.md (adapters + Telegram channels catalog PRs). Agent prompt & storefront: ../README.md.

Telegram channels: shared list telegram-channels.example.json — how to add via PR: TELEGRAM-CHANNELS.md.

What the agent remembers, and how to read it

All local state is in the SQLite file below. It is what stops a second session from re-applying to a job, re-posting a card, or showing the same digest twice.

One call before acting on a card:

workix_job_state {"job_id": "1ca5ff1099b3ca91"}   # or {"url": "https://…"}

It returns whether the card was shown in a digest and when, whether it is already on workix.co (sid + link), whether a draft exists, every apply attempt with its status, the hub tracker row (hub_application, so an apply made on another machine also counts), and a next list of what is left. Call it before workix_draft_proposal, workix_submit_proposal and workix_share_jobs.

Applications live on the hub too. The local outreach table only knows this machine. workix_track_apply records the application on workix.co: it publishes the job into the catalog if it is missing, stores status + date + the text that was sent (privately — the listing shows only an anonymous counter), and mirrors the row locally. workix_submit_proposal calls it automatically when the agent sends the proposal itself. When the human applied by hand, ask them for the text they sent and pass it as text with via: "user" — their own wording is the best base for the next proposal, which is what workix_list_applies is for. workix_sync_applies pulls the history back into a fresh local store, and workix_delete_apply (needs confirm: true) removes a row that should not be there — the mirrored job stays in the catalog, since that listing is public board content rather than part of your private log.

Mirroring is not applying. workix_share_jobs writes an outreach row with channel: "hub" so the mirror is auditable, but those rows are catalog copies, not messages to a client. workix_job_state keeps them out of outreach and out of applied, and reports them under hub_share instead — otherwise a reposted card would look like one you had already answered.

Outreach statuses: draft (written, not sent) · sent · ok (confirmed delivered/accepted) · reply (client answered) · skip (decided against it) · blocked (the platform refused — KYC, connects, ban). Log draft when the text is generated, then update the same id to sent/ok once the user approves and it actually goes out.

Session shape: open with workix_checkpoint_get so search resumes instead of restarting; close (or pause, or switch platform) with workix_checkpoint_set carrying what was done, what is next, which surfaces are finished, and what is blocked.

Local store and shared cache

Everything local lives in one SQLite file, $WORKIX_MCP_DATA/workix.db (default mcp/data/): job cards, drafts, the outreach log, checkpoints, hub shares, and the board response cache. It uses node:sqlite, which ships with Node — nothing compiles at install time, but it does mean Node 22.5+.

An existing store.json is imported on first start and renamed to store.json.migrated; the import runs once and the original is left on disk in case you want it back.

Why not the old JSON file. Every read parsed the whole store, so a 13MB file cost ~126ms per call — and wasShownInDigest() sits inside a filter loop, which turned a digest into a minute of pure parsing. It was also unsafe with more than one session: each MCP process did read-modify-write on the same file, so concurrent writes were lost, and a torn write made the loader fall back to an empty store and discard everything on the next save. Measured after the move: 500 wasShownInDigest calls went from ~63s to 9ms.

**The cache is what makes a second se