Kanwas
Kanwas is an open-source multiplayer canvas where product teams and an AI agent share the same documents, evidence and decisions. Context compounds board after board, and every file stays Git-backed markdown you can take away.
What is Kanwas?
Kanwas is a multiplayer workspace in which a team and an AI agent work on the same documents, evidence and decisions. Rather than separating the chat window from the document and from the file system, it puts a spatial canvas, rich collaborative documents and an agent that can both reason and execute on a single surface. The agent’s tool calls stream as events into the timeline everyone sees, so colleagues watch the work happen instead of receiving only a finished answer. The product rests on an explicit thesis, argued at length on the home page in a piece that takes about fifteen minutes to read: large language models excel at convergent work such as code and fail at divergent work such as strategy because they lack taste; taste is compressed specific context; training averages that context away and destroys it. The conclusion Kanwas draws is that the answer is not more training but accumulating context beside the model, hence its promise of superhuman reasoning plus your specific context. In practice, boards hold code, documents, tasks, embeds and iframes side by side, and those boards, notes, tasks and decisions form a context graph that compounds over time. The agent is configured by the team’s own rules, workflows and skills, and the underlying model is a choice between Claude, GPT and Gemini. The publisher describes a terminal-grade agent that forces a terminal on nobody: work starts in the browser with no download. Documents are markdown files backed by Git with full version history, real-time collaboration, board sharing and access permissions are built in, and the publisher claims more than a thousand connections to team tools plus a command-line tool. Technically, canonical state is a CRDT built on Yjs and BlockNote, the backend is AdonisJS over PostgreSQL and Redis, and agent execution is sandboxed on E2B in the cloud or Docker locally. The whole thing is open source under Apache 2.0, with the kanwas-ai/kanwas repository carrying 747 stars and 102 forks on 7 September 2026, and it can be self-hosted with Docker Compose. Access to the hosted service is gated by a waitlist, and the home page claims a number-one product of the day badge on Product Hunt.
What it does
- Create, edit and share product context on a canvas the whole team works on
- Put an AI agent to work on that same context, with its tool calls visible in everyone’s timeline
- Produce structured deliverables: PRDs, discovery readouts, implementation plans, launch plans
- Connect Slack, Linear, Notion and the codebase to feed the context
- Sync a workspace with the local disk through the CLI (kanwas pull / kanwas push)
- Bulk-import existing markdown with kanwas import ./notes
- Self-host the whole stack with Docker Compose
When to use Kanwas / When not to
A quick filter to help you decide if Kanwas is the right fit.
When to use Kanwas
- Product managers who want interview snippets, tickets and competitor screenshots in one place, then a discovery readout or a PRD that traces back to those sources
- Startup founders juggling a fundraising deck, customer interviews, an MVP spec and a hiring plan on a single canvas
- Engineers who need the spec, the design and the research in one context, produce an implementation plan with tasks and acceptance criteria, then pull the markdown into the repo
- Marketing and sales teams building launch plans, comparing copy variants side by side, or keeping reusable account boards with research, exchange history and stakeholder maps
- Technical teams that refuse vendor lock-in and can self-host with Docker Compose and their own model API key
When not to use Kanwas
- Buyers working under a compliance, legal or procurement gate: no terms of service, no privacy policy, no DPA and no security certification are published
- Teams that need mobile access, since there is no iOS and no Android application
- Anyone who needs hosted access immediately: the waitlist was full on 7 September 2026, with the site stating that the team is currently at capacity
- Developers looking for a public API to build on, as none is documented for end users
- Non-technical teams with no Docker skills and no paid model API key, for whom the self-hosted route is out of reach
How to use Kanwas
A typical end-to-end flow, from setup to results.
- Choose between the hosted service on kanwas.ai and a self-hosted deployment; the hosted route is the quicker one whenever access is open
- For the hosted route, request access through the waitlist form (name, company URL, role, number of PMs in the company — None, 1-5, 5-10, 10-20 or 20+ — and email); the list was full on 7 September 2026
- As an alternative, use the routes the PRD template page suggests: message Johan on LinkedIn to get access, or book a call with the team through Calendly
- Once inside, work straight from the browser: the publisher promises no download and no terminal, just a canvas and an agentic thinking partner
- To self-host instead, install Docker and Docker Compose and obtain an Anthropic and/or OpenAI API key
- Run git clone https://github.com/kanwas-ai/kanwas.git, then cd kanwas, copy the four .env.example files, run docker-compose --profile app up and open http://localhost:5173
- Install the command-line tool with npm install -g @kanwas/cli, run kanwas login and authorise the session in the browser tab that opens; the configuration is stored in ~/.kanwas/config.json
- Bring existing notes in with kanwas import ./notes; only .md files are taken and the folder tree is preserved (options --name, --id, --dest, --overwrite)
- Edit locally with kanwas pull, which opens an interactive picker, then send changes back with kanwas push; the directory is bound to its workspace through a .kanwas.json file
- Day to day, drop the context onto a board, let the agent work on it while the team follows its tool calls in the timeline, collect the deliverable, and script the whole thing in CI with kanwas workspaces --json and the --id or --name flags
Pros & Cons
Pros
- Apache 2.0 open source: the code is auditable, self-hosting costs no licence fee and there is no vendor lock-in
- Documents are markdown files backed by Git with full history, so leaving is trivial and any editor can read them
- Model-agnostic, with a choice between Claude, GPT and Gemini rather than a dependency on a single AI provider
- The agent is observable: its tool calls appear in a timeline the whole team sees, not just the final answer
- One surface for canvas, documents and agent, where the usual workflow is scattered across chat, docs and editor
- A scriptable CLI usable from continuous integration or from another agent
- Unusual signals for a young product: a thesis argued from more than 25 academic references, a named customer testimonial (Samuel Beek, founder of Schematik), a number-one day on Product Hunt, and 747 stars and 102 forks on 7 September 2026
Cons
- No terms of service, no privacy policy, no DPA and no legal notice are published anywhere
- No company name and no postal address: no legal entity can be identified, and the contributor agreement itself refers to a company still to be formed
- No contact email address at all; the only human channel on display is a personal LinkedIn profile, a Calendly link and a community Slack
- Hosted access is closed behind a waitlist that was full on 7 September 2026, and no price is published, so the cost of the hosted version cannot be anticipated
- No security certification, no declared hosting country, no subprocessor list, and no published position on training models with customer data
- No mobile application and no documented public API
- A very young project — domain registered on 30 November 2025, first Wayback capture on 6 May 2026, two public commits — whose self-hosted route also demands Docker skills and model API keys the user pays for
Pricing & Plans
No pricing page exists: /pricing and /plans return 404, and no amount, currency or tier appears anywhere on the site. A lowest price point therefore cannot be stated, because none is published. The source code is released under the Apache 2.0 licence, so a self-hosted deployment carries no licence fee, and the official repository invites readers to “Try Kanwas for free at kanwas.ai”. Access to the hosted version is regulated by a waitlist rather than by payment, and no time-limited free trial is announced. The only residual costs of self-hosting are the model API keys, from Anthropic and/or OpenAI, and the infrastructure, both borne by the user.
Data, GDPR & hosting
A consolidated view of how Kanwas handles your data.
GDPR overview
There is no GDPR implementation to report. The words GDPR and RGPD appear nowhere on the site or in the public repository, as neither a claim nor a denial. No privacy policy, no terms of service, no data processing agreement and no subprocessor list are published; no Article 27 representative is designated and no data protection officer is named. No email address, postal address or company name is published either, so no controller can be identified and no data subject request could be addressed to anyone. The waitlist form collects name, company URL, role, product team size and email with no processing information attached. The site’s only regulatory trace is an Article 4 rights reservation under EU Directive 2019/790 in robots.txt, which concerns copyright in the site content, not personal data. Self-hosting makes the user the controller of their own instance.
Who owns the data?
No terms of service and no privacy policy are published on kanwas.ai: /terms, /privacy and /legal all return 404, and the sitemap declares only three URLs. Ownership is therefore asserted in product copy alone, never in a contract. The official repository states “Your files, your repo. Git-backed markdown filesystem with full version history. No vendor lock-in”, and the home page adds “No lock-in. Your files are yours, with a transparent filesystem under the hood.” Under a self-hosted Apache 2.0 deployment the PostgreSQL database and the markdown files never leave the user’s own infrastructure, so the question settles itself. For the hosted version, no ownership clause exists and nothing is contractually enforceable.
Reuse rights
There are no terms and conditions to read: the publisher has released no contract governing what it does with user data, and it takes no position, in either direction, on whether customer content trains models. What is documented is technical. Canonical state is held as a CRDT and projected into the interface, into backend workflows and into an isolated agent filesystem; the agent reads the workspace — “Underneath, AI agents read your workspace. They connect to Slack, Linear, Notion, your codebase” — and runs in a sandbox, E2B in the cloud or Docker locally. Self-hosted, third-party models are called with API keys the user supplies and pays for. On the user’s side, reuse is unconstrained in practice: documents are markdown files backed by Git, and the CLI command kanwas pull brings them back to disk in full. The robots.txt file carries search, ai-input and ai-train content signals plus a rights reservation under Article 4 of EU Directive 2019/790, but that covers copyright in the website’s own content, not user data.
Data retention & training
Hosting summary
No hosting country and no hosting region is declared for the hosted version of Kanwas. There is no privacy policy, no terms of service and no subprocessor list in which such information could appear, so the jurisdiction governing data at rest is simply unknown. The only observable fact is at network level: kanwas.ai is served from behind Cloudflare (AS13335, anycast address 172.67.68.26), which describes the delivery layer and says nothing about where application data is stored. The published architecture names PostgreSQL and Redis as the data stores and an execution sandbox running on E2B in the cloud or Docker locally, without naming a region for either. Under self-hosting the question resolves itself: PostgreSQL, Redis and the markdown filesystem run in the user’s own containers, on infrastructure the user chooses, under whatever jurisdiction that infrastructure sits in. For the hosted service, nothing allows a buyer to determine where their data would reside, and no certification is displayed to offset that silence.
Things to keep in mind
Risks and trade-offs to weigh before adopting Kanwas.
- Nothing to agree to and nothing to invoke: with no terms of service, no privacy policy, no DPA and no legal notice, a disagreement has no contractual ground to stand on
- No identified legal entity — the contributor agreement points to a company to be formed to steward the project, in the future tense — so there is nobody to hold accountable
- No institutional support channel: no email address, no form, no help centre, only a personal LinkedIn profile, a Calendly link and a community Slack
- Hosted access depends on a waitlist that nobody guarantees will open, which makes a fragile dependency for a team process built on it
- No published price: what a paid tier would cost, and when it might appear, is a complete unknown for a team building habits on the free path
- The publisher takes no position on whether customer data trains models, declares no hosting country and displays no certification, so sensitive material deserves real caution
- Delegating thinking to a shared agent can quietly erode a team’s own reasoning; the tool’s own thesis holds that taste comes from human context, which is worth remembering when a streaming timeline makes the agent look authoritative
Setup & Integrations
Technical difficulty
Two routes, two levels of difficulty. The hosted route claims no setup at all — no download, no terminal — but access depends on a waitlist that was full on 7 September 2026. The self-hosted route is for technical users: Docker and Docker Compose, four .env.example files to fill in, a model API key, then docker-compose --profile app up and http://localhost:5173. The CLI adds a global npm install and a browser-based OAuth login. A public architecture document helps, and local development is documented service by service, but Docker literacy remains a hard prerequisite.
Deployment
Integrations
Behind Kanwas
Fundraising
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Kanwas.
Frequently asked questions
What is Kanwas?
How much does Kanwas cost?
Can anyone sign up freely?
Is Kanwas open source?
Can Kanwas be self-hosted?
Which AI models can be used?
Which integrations are available?
Can I get my files back out?
Is there a mobile app?
Are there terms of service or a privacy policy?
Should you pick Kanwas?
Kanwas makes an unusually clear bet: the product is the shared context, not the chat window and not the storage. A team and an agent work on the same boards, the agent’s steps are visible to everyone, and the output is markdown files backed by Git that nobody can lock away. The thesis behind that choice is argued in public and at length, and the code is open under Apache 2.0. For a product less than a year old, those are two rare signals of seriousness. The counterpart is severe and deserves plain words. There is no contractual footing whatsoever: no terms of service, no privacy policy, no data processing agreement, no legal notice, no company name, no postal address, not even an email address. No hosting country is declared, no certification is displayed, and the publisher says nothing about whether customer data reaches model training. Hosted access sits behind a waitlist that was full on 7 September 2026, and no price is published, so nobody can plan for what a hosted seat will cost. Self-hosting is therefore the one route available without conditions. It costs no licence fee, it keeps the data on infrastructure the team controls, and it demands Docker literacy plus model API keys the team pays for. For an engineering-led team, that is a reasonable experiment. For anyone buying under a compliance or procurement constraint, the documentary void alone is disqualifying today — not because something worrying was found, but because nothing at all has been published to examine. Kanwas is worth watching closely. A domain registered in November 2025 and a first web capture in May 2026 describe a product still taking shape, and this assessment should be revisited often.
- Choosing a selection results in a full page refresh.
- Opens in a new window.