Lanes Link gives every agent one self-hosted endpoint for accounts and memory
Lanes Link is an Apache-2.0 MCP endpoint you run yourself that connects to over 100 providers, limits each agent to a named Access Profile, keeps memory and skills across chat clients, and records every call for later audit.
Image: Lanes
Why it mattersA team that keeps switching AI clients can move accounts, credentials and permissions to one endpoint under its own control, so an agent change no longer means reconnecting Gmail, Slack and GitHub again.
Lanes published a Show HN this week for Personal Context MCP, a hosted-or-self-hosted MCP endpoint called Lanes Link that gives every AI client the same set of accounts, memory, skills and permissions. The post reached 25 points on Hacker News, and the code is Apache-2.0. The argument is that the connectors built into Claude, ChatGPT, Codex or OpenCode belong to each client, so switching agents means reconnecting Gmail, Google Drive, Slack, Notion, Linear, GitHub, Stripe, Sentry, HubSpot and Atlassian all over again. Register one MCP endpoint and every client reaches the same accounts.
One endpoint, many clients
Lanes Link connects to over 100 providers through OAuth and exposes them behind a single Model Context Protocol endpoint. Requests go from the user's machine straight to the provider, and the response comes back to the machine. Lanes says the one exception is the initial OAuth handshake, which needs a client secret it holds; the docs describe lanes link connect gmail --own-client for teams that would rather register their own OAuth client and remove that step too. Lanes says it does not run a model and does not receive the tool output.
Access Profiles instead of one shared set of tokens
Each agent is linked to a named Access Profile that lists which accounts it may reach, which actions it may call, and which team members may use it. The example profiles in the showcase page are a personal profile with four connections, a forward-deployment profile with fifteen, and a support-agents profile with five. The endpoint enforces the profile at dispatch time, so a denied action is refused by the server no matter what the model was prompted to do. An account not granted to a profile is not listed to that profile at all. The stated example is a profile that carries gmail.search = allow and gmail.send = deny regardless of what the agent was prompted to do.
Memory, skills, entities, vault, audit
The endpoint bundles more than credentials. Memory holds facts between sessions; Tasks tracks work items with status; Assets stores files retrievable by name; Skills lets the team supply named procedures the agent must follow; Identity signs an agent's output as the user; Entities looks up people and companies from records instead of the model's memory; Vault releases a credential to the tool without revealing it to the model; Audit records every call, allowed or refused. That last piece is the one a security review will focus on.
The consequence for a team building software is straightforward. If accounts, memory and permissions live in the AI client, every new client is a new migration and every agent decides its own permissions. If they live behind one endpoint the team runs, the agent can be replaced easily and the audit log becomes the record of what actually happened. Lanes Link is free, Apache-2.0, and self-hostable, with a paid managed cloud option if a team would rather Lanes run the endpoint. The trade-off is one more piece of local infrastructure to maintain, plus registering an OAuth client per provider for teams that want to remove the Lanes-held client secret from the process.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
