Agent-native Built on MCP, from the ground up

Your agents update it. You visualise it.

PlanAct is the shared surface between your agentic system and your team. Claude writes topics, boards and tasks straight into it over MCP — you open documents, slides and boards that are already current.

Open PlanAct See how MCP works

Self-hosted on your own Postgres Semantic search runs locally Your content never leaves your box

app.planact.ai/documents

Document Slides Board

Q3 platform strategy

Product Store · Documents Updated 2 hours ago

One store. Written by machines, read by people.

Most tools make you choose: a database your agents can write to, or a workspace your team can actually read. PlanAct is both ends of the same object.

Your agents write

Claude Code, a scheduled job, or your own service calls the MCP tools — creating a topic, moving a card, closing a task.

create_topic · update_task

PlanAct stores it

One Postgres row, embedded with a local model and indexed with pgvector — versioned, permissioned, searchable by meaning.

postgres · pgvector

You visualise it

The same topic opens as a document to read, slides to present, or a board to plan against. Nothing to sync, nothing to refresh.

document · slides · board

The return arrow matters as much as the first one — what you write is what the agent reads next.

The visualise half

The agent wrote one topic. You get three ways to see it.

A topic isn't locked to a format. The same content your agents maintain renders as a document to read, slides to present, or a board to plan against — grouped into stores your team defines.

Document

Long-form writing with structure that survives — the canonical version of what your team decided and why.

Slides

The same topic as a deck when you have to present it — and export to PDF when someone asks for the file.

Board

Columns and cards when the topic is work in motion, with projects and tasks alongside the thinking behind them.

The update half · MCP

One HTTPS endpoint. Paste it into Claude and you're connected.

PlanAct speaks Model Context Protocol over a public HTTPS URL — so Claude Code connects with a single command. No tunnel, no bridge process, no export step. The same topics, the same search, the same permissions your team sees.

Your MCP endpoint https://api.planact.ai/mcp HTTPS · token-authenticated

bash

# point Claude Code at your PlanAct workspace
$ claude mcp add --transport http planact \
    https://api.planact.ai/mcp \
    --header "Authorization: Bearer $TOKEN"

# that's it — the tools show up in the next session
$ claude
> what did we decide about billing in Q3?

  search_topics(q: "billing decision Q3")
  → 3 topics · ranked by meaning, not keywords

Scoped to the caller. A real session token binds the connection to that user's workspace and honours their org memberships — the assistant sees exactly what the person sees.
Read and write. Claude can create a topic, update a board, or file a task — not just retrieve. The work lands in the same place your team already looks.
Two transports, one tool set. HTTP over TLS for Claude Code and claude.ai connectors, stdio for a local process. Identical behaviour either way.
OAuth connectors, next. Per-user consent, scoped grants and revocation — so claude.ai connects to the same endpoint without anyone handling a shared token.

search_topics get_topic create_topic_document create_topic_slides create_topic_board update_topic list_projects create_task update_task

Semantic search

Ask what you meant, not what you typed.

Every topic is embedded with a local model and indexed in Postgres with pgvector. Search matches on meaning, so the answer surfaces even when nobody used your words.

Runs on your box. Embeddings are computed locally — no third-party API sees your content.
Degrades gracefully. If the model can't load, search falls back to keyword matching instead of failing.
One index, two surfaces. The UI and Claude query exactly the same search.

how do we handle refunds after 30 days

0.91

Returns & goodwill policy

"…beyond the standard window, support can approve a credit at their discretion…"

0.84

Billing escalation runbook

"…when a customer disputes a charge outside the refund period…"

0.79

Q2 support retro

"…the 30-day cutoff caused the most friction of anything we shipped…"

Built for organisations

Roles that match how teams are actually shaped.

Four levels, enforced everywhere — the REST API, the UI, and every MCP call. Nothing is bolted on after the fact.

Platform admin

Creates organisations and their first admins. One account, held by you.

Organisation

The tenant boundary. Stores, topics, projects and members all live inside one.

Org admin

Adds people with a password — no email round-trip needed to get a team started.

Member

Reads and writes within their org, through the UI or their own MCP connection.

Stop reading your agents' output. Start seeing it.

Let your agentic system keep the knowledge current — and open a document, a deck or a board whenever you want to know where things stand. On infrastructure you control.

Open PlanAct Take the tour