Put your AI work on the shelf.

One place for everything people make with AI agents — Claude, ChatGPT, Claude Code, Codex.
An agent puts the work on Полка ("the shelf"): it opens by link, the team discusses it, another agent picks it up.

polochka.app · Catalogue · Developer docs · Docs (Russian) · For companies · Русский

- **A shelf for your agents' work.** A page, a prototype, a dashboard or a whole project folder — the agent saves it in one sentence. It opens by link without an account and runs in a sandbox; every edit is a new immutable version; another chat or agent finds the work by its text and continues. - **Memory of agent sessions.** On your own installation, Claude Code and Codex sessions go to the shelf: what the agent did, which commands it ran, which pull request it led to. Secrets are redacted on the machine ([AGENT_SESSIONS](docs/specs/AGENT_SESSIONS.md), Russian). - **For companies.** Self-hosted on one VM, department shelves with roles, sign-in through your IdP. The commercial edition adds an agent control centre for security teams: secrets, addresses and MCP servers, dangerous commands, spend ([COMMERCIAL.md](COMMERCIAL.md)). The interface and most documentation are in Russian. Code, identifiers, commands, API fields and contributor docs are in English. ## Get started Your agent saves the work to Полка. Connect it once; from then on say "save this to Полка" in the chat and the reply contains a link. **Claude** (claude.ai and Claude Desktop): Settings → Connectors → **Add custom connector**, URL ```text https://polochka.app/mcp ``` → Add → Connect → Allow in Полка. **Claude Code**: the plugin installs the MCP server and the Полка skills in one command. ```sh claude plugin marketplace add artkruglov/polka-plugin && claude plugin install polka@polka ``` Then in Claude Code: `/mcp` → `plugin:polka:polka` → Authenticate. **Codex**: the same plugin for Codex. ```sh codex plugin marketplace add artkruglov/polka-plugin && codex plugin add polka@polka codex mcp login polka ``` **Another MCP client** (Cursor, Gemini CLI, Windsurf…): the remote server `https://polochka.app/mcp` (Streamable HTTP, OAuth sign-in) plus the skill `npx skills add artkruglov/polka-plugin`. Each time Полка opens: sign in to your shelf (or start without signing up) and press Allow; no token or password passes through the agent. Or tell a terminal agent: `Connect Полка: https://polochka.app/connect`. You can still upload a file from your computer without an agent. Details: [connecting agents](docs/connect-agents.md) (Russian). ## Features

The agent saves it for you In Claude (claude.ai and Desktop), Полка is a connector: say "save this to Полка" and the reply contains a link. ChatGPT connects to the same address, but saving from it hasn't been checked yet. Claude Code and Codex install the plugin with one command and no token; other MCP clients use the `/mcp` address; scripts and CI use the HTTP API.	Interactive for recipients React/JSX chat artifacts are compiled into one self-contained page, with libraries bundled in and no network access. The page opens in a sandbox on a separate domain, `polochka.page`.	Exact versions Every save is immutable and has a SHA-256. A link shows the version you published, not your latest draft.
Revocable links Links last 1, 7 or 30 days, can be revoked at any time, and recipients can report them. Private works stay out of the catalogue and search indexes.	Shelf, folders and search Covers with the heading and the opening lines, full-text search across works, trash. The agent sorts the shelf into folders itself ("sort my shelf") and saves new work into the fitting folder. Team template libraries.	Data stays in Russia polochka.app runs on Yandex Cloud: the database, files, backups and e-mail are stored and processed in Russia (privacy policy, Russian). Or run Полка yourself.
Projects A folder of linked pages — HTML, Markdown and React components — becomes one work: a file tree, links between pages that work, and one link for the whole project.	Department shelves A department gets its own shelf: members and roles, the agent picks the shelf when it connects, links and discussions, a company admin page. Turned on with `TEAM_SHELVES=on`. On polochka.app since 12.10.2026: each company is its own organisation and its company admin sees only its shelves. To get a department shelf in the cloud, write to hello@polochka.app and we will appoint your company admin.	Calm moderation Rules and models check every link. A prototype with a sign-in screen or a research page with hundreds of sources isn't held; recipients see a warning only when a page asks for a password, a code or card details.

## What it looks like

_{The recipient. An interactive page runs in a sandbox, no account needed.}	_{The shelf. Folders, covers with the heading and opening lines, search and sorting across the whole shelf.}
_{Лента (Feed). Interactive pieces from the Полка editors.}	_{For companies. Cloud, self-hosting or a commercial license.}

## How it works ```mermaid flowchart LR A["Agent
Claude.ai · ChatGPT
Claude Code · Codex"] -- "MCP /mcp
(OAuth 2.1 or token)" --> P S["Script, CI"] -- "POST /api/v1/publish" --> P U["You, in the browser"] -- "upload, paste code" --> P P["Полка · polochka.app
versions, links, templates"] --> DB[("PostgreSQL")] P --> S3[("Versioned S3")] P -- "link /s#…" --> R["Recipient"] R -. "iframe, sandbox without network" .-> V["Viewer · polochka.page"] ``` The agent hands over the work's code itself; Полка doesn't pull anything out of the chat. Every save becomes an immutable version, and a link is bound to a version. Foreign HTML is treated as hostile: an interactive page opens on a separate domain, with no access to cookies, Полка's API or the network. More: [architecture](docs/en/architecture.md), [threat model](SECURITY.md#модель-угроз-вкратце). ## Five ways to save a work | From | How | Details | |---|---|---| | Claude (claude.ai, Claude Desktop), ChatGPT (being checked) | Connector `https://polochka.app/mcp` with OAuth 2.1 sign-in. The model calls `polka_publish` and returns a link | [Connector](docs/MCP_CONNECTOR.md) | | Claude Code, Codex | The Полка plugin in one command: MCP server and skills. Sign in and allow in the browser; no token | [Connecting agents](docs/connect-agents.md#плагин-полки-для-claude-code-и-codex) | | Other MCP clients | Token from the «Агенты» (Agents) page, Streamable HTTP at `/mcp` | [Connecting agents](docs/connect-agents.md) | | Scripts, CI, in-house agents | `POST /api/v1/publish` or the dependency-free CLI `scripts/polka-publish.mjs` | [HTTP API](docs/PUBLISH_API.md) | | By hand | Upload a file (HTML, text, PNG/JPEG/WebP up to 5 MB), or paste code on the «Сохранить» (Save) page | [FAQ](docs/faq.md) | Saving and the link were checked by hand with Claude.ai, not yet with ChatGPT ([status](docs/status.md)). There is no saving by a link to a Claude or ChatGPT artifact: Полка's server can't fetch it (the sites require a login and sit behind Cloudflare), so the agent hands over the work instead. **For agent developers:** the Claude Code and Codex plugin (`.claude-plugin/`, `.codex-plugin/`, `.mcp.json`, `skills/`) lives here and installs from its light copy [artkruglov/polka-plugin](https://github.com/artkruglov/polka-plugin), built by `scripts/plugin-repo.mjs`; the skill alone installs with `npx skills add artkruglov/polka-plugin`; agent reference at [/llms.txt](https://polochka.app/llms.txt), HTTP API at [/openapi.json](https://polochka.app/openapi.json). ## A workspace for agents A project work is a folder with versions: an agent takes it, edits it and saves the next version, and people open it by link without an account. See [POSITIONING](docs/specs/POSITIONING.md) (Russian), [AGENT_WORKSPACE](docs/specs/AGENT_WORKSPACE.md) and [PUBLISH_API](docs/PUBLISH_API.md). - **The folder over MCP, no network needed:** `polka_list_files`, `polka_read_file`, `polka_change_files` (add, replace, remove files; the server copies the rest); over HTTP `GET /api/v1/works/:id/file`, `POST /api/v1/works/:id/changes`; from a terminal `polka pull/push`. - **Accepted version and an owner** per work (the work menu, «Принятая версия»). Marking a version does not move any link, and a new version does not clear the mark. - **Agents read the shelf:** `GET /api/v1/works` (text search with ranking, `since` for what changed), `GET /api/v1/events` (a change feed to poll), `polka_list` with the same parameters; the shelf's «how we do things here» note arrives in `polka_context`. - **Link mode:** an unattended agent (a service account) does not move a pinned link until a curator switches it to follow new versions; an agent with your own token works as before. - **On polochka.app and on your own installation** (`TEAM_SHELVES=on`, `SERVICE_ACCOUNTS=on`): search across department shelves, service accounts with a responsible person, and short task tokens for cron and CI. ## Quick start You need Node.js ≥ 22.16, npm and a running Docker. ```bash git clone https://github.com/artkruglov/polka.git && cd polka npm ci npm run local:setup # .env with unique secrets and the local interactive viewer (HTML_LIVE_MODE=local) npm run infra:up # PostgreSQL 16 + MinIO, bound to 127.0.0.1; the first run builds MinIO from source, ~10 min npm run db:migrate npm run storage:bootstrap-local # versioned bucket and a capability check npm run admin -- account create demo --generate # login and password in .local/demo-account.txt npm run build npm run dev # http://127.0.0.1:4390 ``` `local:setup` prints the same sequence at the end. If the default ports are taken (or this is a second clone), choose your own when `.env` is created: `POLKA_LOCAL_PROJECT=polka-two POLKA_LOCAL_PG_PORT=55432 POLKA_LOCAL_S3_PORT=9138 PORT=4490 VIEWER_PORT=4491 npm run local:setup`. > [!WARNING] > If `db:migrate` fails with an authentication error, volumes from an earlier install with an old password are still there. Reset, deleting the local data: `docker compose --env-file=.env -f deploy/compose.local.yml down -v`, then start again from `npm run infra:up` ([details](docs/en/local-development.md)). The interactive viewer, e-mail code sign-in and the separate test suites are covered in [docs/en/local-development.md](docs/en/local-development.md).

Checks

```bash npm run check # frontend layers + TypeScript npm run build npm test # throwaway database and bucket, removed afterwards npm test -- --live # suites that need the local viewer npm run verify # everything before a push; GitHub Actions runs the same on every pull request ```

## Self-hosting A deployment is one Docker image plus external PostgreSQL and versioned S3 storage. You build the image from source (`docker build`); no published image exists yet. The interactive viewer must run on a separate registrable domain. The recommended path is [deploy/hosted/README.md](deploy/hosted/README.md): a single VM behind Caddy, which is how polochka.app runs. [deploy/BASE.md](deploy/BASE.md), [deploy/RESTORE.md](deploy/RESTORE.md) and [deploy/VIEWER_STAGING.md](deploy/VIEWER_STAGING.md) are drafts for experienced operators. For how the parts fit together, see [docs/en/architecture.md](docs/en/architecture.md). ## For companies | | polochka.app cloud | Self-hosted (open core) | Commercial edition (not in this repository) | |---|---|---|---| | Price | Free during the pilot | Free under the AGPL-3.0 | By agreement | | Where the data lives | Yandex Cloud, Russia | Your servers | Your servers | | What's included | The open core: department shelves with roles (we appoint a company admin on request to hello@polochka.app), service accounts for cron and CI, webhooks | The whole open core: department shelves, roles, agents per shelf, company admin, sign-in through an IdP over OpenID Connect, webhooks | The core plus a closed extension for organisations: an agent control centre (employees' sessions, secrets, addresses and MCP, dangerous commands, spend), employee-only links, a link policy, an agent limited to a folder, an agents' journal exported to a SIEM, Jira and Bitrix24 comments, export of accepted versions, SAML sign-in and an employee directory over SCIM; next — offline installation | | Your code changes | — | If people use your modified Полка, publish them under the AGPL-3.0 | May stay private | | Support and SLA | — | — | By contract | More on the [For companies](https://polochka.app/enterprise) page and in [COMMERCIAL.md](COMMERCIAL.md). Sign-in through a company IdP (OpenID Connect), Yandex ID, VK ID and Google, and template-library access by e-mail domain already work in the core. SAML and SCIM are in the commercial edition (since 0.9.0); there are no team accounts yet ([roadmap](docs/roadmap.md)). ## Limitations - **Pages are capped at 5 MB and have no network.** Everything must be inside one HTML page or bundle, with styles, images and fonts as `data:` URIs. External scripts, `fetch` and forms don't work for recipients. - **Interactive mode needs a separate viewer domain.** Without one (`HTML_LIVE_MODE=disabled`), recipients see a static page with scripts off. - **Claude/ChatGPT links can't be imported.** Save through the connector, download the file or paste the code. - **Downloaded copies can't be revoked.** Revoking a link closes it, but it can't delete what a recipient already downloaded. - **Sign-in is by e-mailed code** (SMTP required), by an operator-issued password, or through Yandex ID, VK ID, Google or a company IdP over OpenID Connect ([SIGN_IN_PROVIDERS](docs/specs/SIGN_IN_PROVIDERS.md), Russian). Sign-up can be open, limited to listed addresses or mail domains, or capped per day. SAML and SCIM aren't in the open core; they're in the commercial edition. - **URL import** (`URL_IMPORT_ENABLED`, experimental, server and MCP only; the web app no longer offers it) and **account deletion** (`ACCOUNT_DELETION_ENABLED`) are off by default. More: [docs/faq.md](docs/faq.md) (Russian). ## Status > [!NOTE] > **Current release — `v0.19.0`** ([CHANGELOG](CHANGELOG.md)). A hosted pilot runs at https://polochka.app; e-mail sign-up is open to any address, up to 50 new shelves a day. The API, database schema and UI may still change. What works and what doesn't: [docs/status.md](docs/status.md) (Russian). Next, per the [roadmap](docs/roadmap.md): running the pilot (a restore drill, an upgrade guide), publishing the browser extension and checking the ChatGPT connector, then variants of a work, a shelf snapshot by date and `polka pull/push`; later a Telegram bot. ## Documentation In English, for developers: [Developer docs](docs/en/README.md) · [Architecture](docs/en/architecture.md) · [Local development](docs/en/local-development.md) · [Contributing](CONTRIBUTING.md) In Russian: [Document map](docs/README.md) · [Architecture](docs/architecture.md) · [Connecting agents](docs/connect-agents.md) · [Connector](docs/MCP_CONNECTOR.md) · [HTTP API](docs/PUBLISH_API.md) · [FAQ](docs/faq.md) · [Status](docs/status.md) · [Roadmap](docs/roadmap.md) · [Changes](CHANGELOG.md) ## Contributing Issues and pull requests are welcome in Russian or English. How to run the project, which checks to run and the house rules are in [CONTRIBUTING.md](CONTRIBUTING.md). Pull requests are accepted under the [CLA](CLA.md): write "I accept the CLA (CLA.md)" in the description once. Participants follow the [code of conduct](CODE_OF_CONDUCT.md). Where to ask questions: [SUPPORT.md](SUPPORT.md). ## Security Report vulnerabilities privately through [GitHub Security Advisories](https://github.com/artkruglov/polka/security/advisories/new), not in a public issue. The process and the threat model are in [SECURITY.md](SECURITY.md). ## License [GNU AGPL-3.0](LICENSE) or a [commercial license](COMMERCIAL.md) (dual licensing). **From version 0.1.0 on, Полка is AGPL-3.0; releases up to and including v0.1.0-rc.5 remain Apache-2.0**, and that does not change. Running Полка unmodified, including as a service for your team, and modifying it in the open are free under the AGPL-3.0. If you let people use a modified Полка over a network, you must offer them its source: set `SOURCE_URL` to it ([deploy/hosted](deploy/hosted/README.md#исходный-код-изменённой-версии)). A commercial license is for keeping your changes private, embedding Полка in a proprietary product, or getting support and an SLA; see [COMMERCIAL.md](COMMERCIAL.md). See also [NOTICE](NOTICE) and [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md). The project grew out of [Lanka](https://github.com/artkruglov/lanka); see [ACKNOWLEDGEMENTS.md](ACKNOWLEDGEMENTS.md).

Put your AI work on the shelf.

The agent saves it for you

Interactive for recipients

Exact versions

Revocable links

Shelf, folders and search

Data stays in Russia

Projects

Department shelves

Calm moderation