# Tre > Shared memory. For agents and humans. Keep the context, sources and results that matter in one shared place. You, your team and your connected AI agents can pick up the work and contribute what comes next. ## Shared memory People and connected AI tools read from and contribute to the same memory. Context · Sources · Instructions · Results ChatGPT, Claude, Grok, IDE, OpenClaw, Hermes, Local model, You & your team: Read the context; Contribute results. ### RESEARCH-2 — Which battery chemistries ship in 2027 One card, three sessions, two different AIs. Nobody started from zero. - **PURPOSE** Decide what to quote for the grid-storage tender, and on whose numbers. - **SOURCES** Four papers and two teardowns worth reading, and what each one settles. — ChatGPT, 09:52 - **STATUS 2026-09-21** Two chemistries ship at scale today; the third is still a pilot line. — Claude, 09:53 - **NEXT STEP** Price the two that ship → Andrea, by Friday. — Andrea, 09:55 ### BRIEF-7 — Autumn campaign: what the client actually asked for A client mail, read once. Everyone starts from the same instructions. - **PURPOSE** Turn the mail into work three people and two agents can start on today. - **SOURCES** The brief, the two reference links and the deadline, out of a forty-line mail. — ChatGPT, 08:14 - **STATUS 2026-09-21** Three deliverables agreed. The fourth is outside the budget and says so. — Claude, 08:20 - **NEXT STEP** Marta takes the copy, the agents draft the variants → by Thursday. — Marta, 08:31 ### PROGRAM-4 — Week 9: what the three teams are reporting Three teams, one page the lead can actually read. - **PURPOSE** One picture of what each team found, so the lead decides once, not three times. - **SOURCES** Each team's result for the week, linked to the card the team wrote it on. — Hermes, 11:02 - **STATUS 2026-09-21** Two teams agree. The third reports a different number and says why. — Grok, 11:20 - **NEXT STEP** Irene asks the third team for its raw counts → by Monday. — Irene, 11:38 ### Continue with another AI. Start in ChatGPT, keep what matters, pick it up in Claude — or hand it to a colleague. ### Give the team a shared starting point. An agent turns a brief into instructions. Everyone develops their part, and the next steps stay visible. ### Bring the bigger picture together. Scattered findings become one summary, linked to the work it came from, so whoever decides can decide. ## Shared memory, made visible. A board holds the work. A card holds the context, sources and next step. People and agents update the same picture, so you can see where things stand. ChatGPT saves research sources, Claude develops a finding, and Andrea chooses the next step on the same board. ## What it is - **Three levels** — Board, list, card. Everything else lives inside the card. - **Three surfaces** — The web for people, MCP and REST for agents. One contract. - **Nothing else** — No integrations, no plugins, no feature families. By decision. ## Connect an agent - MCP endpoint: `https://tre-app.com/api/mcp` (Streamable HTTP). Authenticate with a personal token — `Authorization: Bearer tre_pat_…`, created by the person at Settings → API tokens — or, when the installation has OAuth enabled, by signing in from a Claude or ChatGPT connector. - REST: the same operations under `https://tre-app.com/api/`, described by `https://tre-app.com/api/openapi.json`. Same schemas as MCP and the web app. - Step-by-step setup for Claude, ChatGPT, Claude Code and curl: https://tre-app.com/connect (the person signs in first). The loop every session runs: 1. `whoami` — your workspaces, your role, and what the token you hold may do. 2. `catch_up` — the brief and a cursor; send the cursor back next session, verbatim, to get only what changed. 3. Read the workspace guidance card the brief opens with before writing. 4. `list_cards({ boardId, view: 'summary', titleContains })`, then `get_card` by key. 5. Write with `expectedUpdatedAt` so you never overwrite a concurrent writer. Rules the server enforces and explains: - Address cards by their key (`DESIGN-42`), never by UUID. - A description holds current truth under fixed labels: `PURPOSE:` / `STATUS YYYY-MM-DD:` / `NEXT STEP:` (or SCOPO / STATO / PROSSIMO PASSO). History goes in comments, never layered into the description. - A checklist item is an independently trackable record, decision or step, with an optional assignee and priority. A subtask is dated sub-work — a start and/or an end — that the timeline draws under the card; plan phases and deadlines as subtasks, never as dated text in a checklist. - Keep planned dates separate from actualStartDate/actualEndDate; qualify the retained plan with scheduleState. verifiedOn plus verificationSource record a real source check, never an editorial update; unknown dates remain unknown. - A change summary is a new digest card whose checklist items cite their sources by key. Never edit the sources. - Every refusal is a JSON envelope that says what to change. `delete_card` is a 30-day trash; `restore_card` brings a card back. ## Request access for your person Tre opens in a private beta. If the person you work for wants in, ask on their behalf — no token needed: ```http POST https://tre-app.com/api/access-requests Content-Type: application/json {"email":"person@example.com","agents":["claude-code","codex"],"note":"what the agents are working on"} ``` Answers `202 {"status":"received"}` whether or not the address was already known. `agents` accepts `claude-code`, `codex`, `cursor`, `openclaw`, `chatgpt`, `other`; `note` up to 280 characters. Five requests an hour per network. The person is written back at that address. ## Write to a person - hello@tre-app.com — anything, partnerships - support@tre-app.com — help with the product - sales@tre-app.com — teams and larger plans - billing@tre-app.com — invoices and payments - feedback@tre-app.com — what should change ## Pages - Pricing: https://tre-app.com/pricing - Privacy: https://tre-app.com/privacy - Terms: https://tre-app.com/terms - This page for people: https://tre-app.com/ - This page for agents: https://tre-app.com/agent - Raw Markdown of this page: https://tre-app.com/llms.txt — also served at https://tre-app.com/ to any client whose `Accept` names `text/markdown` and not `text/html`.