Summary

Pico engine 1.10 is a security release. It lets a remote MCP client like Claude connect to an engine through standard OAuth, with a person approving the client using a passkey and the agent seeing only the channels that person chose to expose.

Pico engine 1.8 gave agents an MCP surface, and 1.9 let picos serve their own views. Both made picos easier to reach, which was the point. But reachability raises an obvious question: who is reaching in, and what are they allowed to do? Pico engine 1.10 addresses that. It's a security release built around letting a remote MCP client, like Claude, safely connect to an engine. A person signs in with a passkey and approves the client, and the client then sees only the channels that person chooses to expose. The major points:

  • Standard MCP authorization. The engine acts as its own OAuth authorization server, with the discovery documents MCP clients expect, PKCE (S256 only), tokens bound to a resource per RFC 8707, Client ID Metadata Documents, and two scopes: mesh:read and mesh:act. A person approves an MCP client through the mesh's existing OAuth ruleset, io.picolabs.oauth, signing in with a passkey.
  • Default-deny MCP exposure. A new MCP policy sits beside each channel's event and query policies. No channel is visible to an agent until someone allows it, and the channel's existing policies still decide which events and queries succeed.
  • Whole-mesh MCP. A mesh endpoint gives an agent pico_tree, pico_channels, query, and raise_event across a subtree of picos, still gated channel by channel.
  • The engine's internal API stays internal. The routes that the developer UI and the local developer MCP server use are now closed to the internet. The engine no longer treats a spoofed Host: localhost header or a proxied request as a local caller; internal calls need a passkey session or the engine's internal token.
  • Secrets stay out of view. Authorization, cookie, and internal-token headers are stripped before rulesets see them. Secret PDS settings return a mask to everyone except the ruleset that owns them, and logs never contain secret values.
  • Webhook signatures. Services like Square sign each webhook so the receiver can tell a real notification from a forged one. Checking that signature takes two things picos didn't have. The signature covers the exact bytes the sender posted, and a pico normally sees only parsed attributes, where a stray space or a different key order breaks the match; event:raw_body() hands a ruleset those original bytes on channels set to keep them. The comparison also has to take the same time whether or not the signature matches, or an attacker can recover it a byte at a time by measuring response times; math:timing_safe_equal() does that. Together with the existing math:hmac(), a pico can now safely accept signals about payments, orders, and other events from outside services.
  • Shorter-lived tokens for public clients. Access tokens last an hour and refresh tokens rotate on every use. Replaying an old refresh token revokes the grant.

Picos are meant to be personal: your data, your rulesets, your mesh. Letting an AI agent act on them is only useful if you stay in charge of what that agent can touch. The changes in 1.10 help ensure that decision stays with the person. An agent gets in because you approved it, sees what you exposed, and does only what each channel's policies allow. Default deny matters here because forgetting to configure something should leave you safe, not exposed. The long-standing principle of use-specific channels protected by policies is now being applied to a new kind of caller. The release notes cover the upgrade steps, and you'll want 1.10.1, which fixes a startup bug in 1.10.0.


Please leave comments using the Hypothes.is sidebar.

Last modified: Wed Nov 18 15:08:11 2026.