Connect Allbase

Bring your knowledge into the agents you already use. Choose your app, approve its access, then check that it can find a note. Each OAuth connection has a name and access you can revoke from Agents. You do not need an Allbase API key.

New here? Sign up / Sign in with an email code. There is no password or separate account creation step. Agent connections become available after private-beta approval.

Connect → retrieve → save → hand off

Connecting authorizes an agent. Retrieving a note proves it can use your knowledge. Saving and checkpointing require write access. Automatic use needs instructions in the agent as well.

Coding agents — one prompt

In Claude Code, Codex, or Cursor, paste this to start setup. Follow the browser sign-in and access review when prompted:

Set up Allbase by following the official instructions at https://allbase.md. Complete the setup yourself, use OAuth, ask before writing global agent instructions, and verify allbase_status when connected.

The agent adds the MCP server, walks you through the browser OAuth grant, and offers the memory instructions below. The connection uses https://allbase.ai/mcp. Already connected? Skip to verification.

Set up manually

These steps configure the connection in an existing installation of your agent. Keep other server entries when editing configuration.

Claude Code

claude mcp add --transport http --scope user allbase https://allbase.ai/mcp

Then open Claude Code, run /mcp, and authenticate Allbase in the browser. Claude Code MCP guide.

Codex CLI

codex mcp add allbase --url https://allbase.ai/mcp
codex mcp login allbase

Start a fresh task if the Allbase tools are missing after setup. Codex MCP guide.

Cursor

Add to Cursor or merge into ~/.cursor/mcp.json:

{ "mcpServers": { "allbase": { "url": "https://allbase.ai/mcp" } } }

Open Cursor's MCP settings and finish the OAuth connection. Cursor MCP guide.

Make Allbase part of the agent's routine

The instruction adapter asks the agent to bootstrap, checkpoint, and cite Allbase during normal work. It is a separate setup step; OAuth alone does not install it. Review the proposed text and approve any change to your global instructions:

claude-code.md → ~/.claude/CLAUDE.md
codex.md → ~/.codex/AGENTS.md
cursor.mdc → ~/.cursor/rules/allbase.mdc

Chat apps

Choose your app below. Your plan and organization settings may affect whether a custom connection can be added.

Claude — web and desktop setup

Go to Customize → Connectors → + → Add custom connector.
Name it Allbase, set the URL to https://allbase.ai/mcp, and add it.
Click Connect: sign in to Allbase, name the agent, choose read-only or read & write, and approve.
Enable the connector in a chat from the tools menu.

For a team account, an owner may need to add the connector before members can connect. Enable it in each conversation where needed. Check mobile availability and account requirements in the Claude connector guide.

ChatGPT — web setup

Go to Settings → Security and login and enable Developer mode.
Open ChatGPT's Plugins page, use the plus button to create a developer-mode app, name it Allbase, set the MCP server URL to https://allbase.ai/mcp, and choose OAuth.
Approve the Allbase consent screen: name the agent and choose its access level.
In a conversation, select Developer mode from the plus menu and enable Allbase.

If these controls are unavailable, check your account and organization settings in the ChatGPT developer-mode guide.

Optional routing instructions

Use the following policy in a supported custom-instructions or skill surface to ask for relevant Allbase retrieval. If your account supports skill uploads, the example is formatted as a SKILL.md. Installation and automatic use still need verification in your account; this does not configure checkpointing.

---
name: allbase-memory
description: Use Allbase whenever a request may depend on the user's durable knowledge, prior project decisions, preferences, project state, agent handoffs, or previously stored context. Trigger on questions like "what did we decide?", "where did we leave off?", "continue this project", references to prior architecture or implementation choices, or ambiguous internal project terms.
---

# Allbase shared memory

Retrieve relevant Allbase context via the Allbase connector before answering
when the request depends on the user's prior knowledge or decisions.

Treat Allbase as authoritative for: prior decisions, user preferences,
project state, architecture choices, completed work, blockers, next actions,
and agent handoffs.

Do not treat Allbase as authoritative for general world knowledge, current
events, public documentation, or software facts that may have changed.

If Allbase returns nothing relevant, answer from other appropriate sources
and say the answer did not come from Allbase — never conclude that no
information exists from an empty Allbase result alone.

The Skill is the routing policy; the connector stays the actual tool. It retrieves only — it never checkpoints conversations on its own.

Any other MCP client

Streamable HTTP at https://allbase.ai/mcp with OAuth. Agent-readable setup instructions live at allbase.md.

Verify the connection, then try a handoff

Check identity and access. Ask the agent to run allbase_status. Confirm the workspace, agent name, and read or read/write access are the ones you intended.
Retrieve real knowledge. Create a small note in Knowledge, then ask the agent to read it and cite its path and revision. A status response alone does not prove that retrieval works.
Check saving separately. If you granted write access, ask the agent to save a small test note, then open it in Allbase. Read-only agents can retrieve but cannot save or checkpoint.
Allbase a small project. Ask the first to checkpoint a small test project with a clear next action. Connect a second named agent to your workspace and ask it to bootstrap that same project key. Compare the returned next action and attribution with the first agent's checkpoint.
Check the routine. In a fresh task, confirm the installed instructions cause the agent to consult Allbase when relevant. An explicit successful test does not prove automatic use.

Use Allbase to run allbase_status and report my workspace, your agent identity, and your effective access. Ask me which existing note to use, then retrieve it and cite its path and revision. Do not write anything for this check. Report any denied or unavailable step separately.

Try saying “Allbase this project”

Once you have verified write access, use this in a project you are already working on with an agent:

Allbase this project. Save a concise checkpoint in my workspace with the current objective, decisions, blockers, and next actions. Use the existing project key, or ask me to name it. Then give me a prompt for my next agent to retrieve that checkpoint and continue.

Review observed activity and refresh connection health in Agents. No recent activity means there is no recent proof; it does not automatically mean the connection is broken.

Open Projects to browse the saved workstream and its checkpoint history. Create a project with that key and invite a teammate: after joining, they and their authorized agents can read the full project checkpoint history. Agents with write access save progress in their own lanes. Link shared notes separately; each file share has its own acceptance.

Something did not work

Tools missing: enable Allbase for the conversation, refresh its tool list, or start a fresh task after configuration.
Sign-in expired or access revoked: reconnect with OAuth. Review the workspace and access before approving.
A shared note is denied: check that its owner shared it with you and that you accepted the share. Joining a project alone does not grant access to every linked note.
Checkpoint denied: check write access; use a read-only agent for retrieval or reconnect with write access if needed.
Agent forgets to use Allbase: review its instruction adapter or routing policy and test again in a fresh task.

Compatibility and observed evidence

Guide reviewed September 4, 2026. The setup routes below are intended capabilities, supported by the linked client documentation. A documented route is not an Allbase end-to-end test. Observed checks below describe the tested surface and their limits; missing observations remain unverified.

Scroll the table sideways to read the evidence column.

Allbase harness evidence as of September 4, 2026
Harness	Intended setup and use	Observed end-to-end evidence
Claude Code	HTTP + OAuth; retrieval; note writes and checkpoints with write access; instruction adapter for routine use.	September 4, 2026, version 2.1.257: an existing authenticated connection ran status and bootstrap in a fresh process and resumed a Codex checkpoint. Fresh OAuth, note retrieval/writes, checkpoint creation, and automatic use were not tested in that run. Plan-only mode blocked MCP calls; normal permission mode with the same read-tool allowlist passed.
Codex desktop app	HTTP + OAuth; retrieval; note writes and checkpoints with write access; instruction adapter for routine use.	September 4, 2026: this review session used an existing connection for status, bootstrap, note reads/writes, and checkpoints. App build was not recorded. Fresh OAuth, automatic use without task instructions, and a first-time user handoff remain unverified.
Codex CLI	HTTP + OAuth; retrieval; permitted writes/checkpoints; instruction adapter for routine use.	Not observed in this review. Desktop-app evidence does not establish CLI setup, version, or routine use.
Cursor desktop	Remote MCP + OAuth; retrieval and permitted writes; Cursor rule adapter.	Not observed in this review. No dated client version, retrieval, write, checkpoint, or automatic-use result recorded.
Cursor cloud	Not evaluated for this release; verify its own configuration and authorization support.	Not observed. Do not assume desktop configuration or access carries over.
Claude web	Custom remote connector + OAuth; retrieval and permitted writes. Routine use depends on client configuration.	Not observed in this review. OAuth, retrieval, writes, checkpoints, and routine use need separate observations.
Claude desktop	Custom connector where available for the account; verify tool enablement in the desktop app.	Not observed in this review. Web setup instructions do not establish desktop behavior.
Claude mobile	Verify the connected account's tool availability in the mobile app.	Not observed in this review. No mobile retrieval, write, or checkpoint result recorded.
ChatGPT web	Developer-mode app + OAuth; retrieval and permitted writes. Optional retrieval policy does not install checkpoint automation.	Not observed in this review. Account eligibility, tool enablement, writes, and checkpoint handoff need separate checks.
ChatGPT desktop	Not evaluated here; verify tool availability separately from the web setup.	Not observed in this review. Web configuration does not establish desktop behavior.
ChatGPT mobile	Not evaluated here; verify tool availability separately from the web setup.	Not observed in this review. Web configuration does not establish mobile behavior.
Other MCP clients	Streamable HTTP + OAuth where the client supports Allbase's authorization flow.	Not observed. Verify the actual client and version before relying on a workflow.

These entries describe documentation evidence, not live status of your account. Check your own connections in Agents. The dated observations above have not been extended by this website update.

Bring a reusable skill into a project

Knowledge → Library supports private skill drafts, review, immutable releases, and exact revision downloads. Choose “Use in a project” on a release to prepare a request for the local helper. MCP setup alone does not install or execute a skill.

Managed installs currently support eligible documentation-only skills in Claude Code repositories. The helper reviews the target and requires approval before local changes. Other harness installs, blueprints, and team distribution are not generally available. See supported installs and verification.