Date: 2026-09-18. The client-side runs used Aside CLI 1.26.906.1630 and background service 1.26.917.1731. This note records the native MCP route only; it does not demote or replace the existing CLI installation route.

Aside exposes MCP configuration at Settings > Plugins & MCPs > MCPs. Per-account configuration is stored at ~/.aside/u/<n>/settings.json, under mcp.servers and mcp.inventories.

Two otherwise identical server registrations were compared:

Registration by itself was therefore not sufficient. Using the GUI's per-server Refresh tools action on the hand-written entry changed it from 0 tools cached to 1 tool cached. A newly created aside exec session then listed mcp__aside-codemode__execute_code and could call it. No daemon restart was needed between the inventory refresh and that new session.

The captured MCP exchange showed:

1. The client identified itself as the Aside browser and sent an initialize request with protocolVersion 2025-11-25.
2. The server's initialize response echoed 2025-11-25.
3. tools/list returned execute_code.

The initialize and discovery exchange completed without a protocol or tool-schema error. This separates inventory attachment from protocol compatibility: the server could complete the handshake even when a manually written registration still had no cached inventory for agent attachment.

When the background service spawned the MCP server, the child received only six environment variables. Its working directory was the background service's own application directory, not the project or account directory. The restricted PATH did not contain ripgrep, so the first attached MCP call failed with ERG404.

The raw environment values, absolute working directory, account identity, and run identifiers are intentionally omitted. They are machine records, not facts needed to judge the result.

The server configuration was changed to pin rgPath to the absolute path of the ripgrep binary bundled with the Aside runtime. That binary reported ripgrep 15.2.0 with PCRE2 support and ran with a completely empty environment, so the pin did not depend on the six-variable child PATH.

After refreshing the server inventory and starting a new aside exec session, the agent used mcp__aside-codemode__execute_code once. That one MCP call ran search.content and returned rows with complete:true, truncated:false, and partial:[] in 40 ms. The agent used no other tool in that run.

The attached execute_code tool description measured 6,775 bytes when the MCP route was first proven. It is resident in every MCP session. For comparison only, the account AGENTS.md block measured 3,808 bytes and the installed user skill measured 8,973 bytes and is loaded on demand.

The same release then moved the per-action synopses out of that description and into the existing on-demand action catalog, leaving the cross-cutting warnings resident. Measured after the change with Buffer.byteLength(TOOL_DEF.description): 1,793 bytes.

Later releases spent some of the remaining budget on warnings that were measured to be worth it: the search envelope, the symlink rule, and the instruction to reach for code mode first. Re-measured the same way at 0.8.0: 2,042 bytes, under the 2,048-byte budget the suite enforces. The READMEs quote this number, and a test recomputes it from TOOL_DEF so the prose cannot drift away from the code again - which it had, at 1,984 bytes.

0.8.1 spends part of that budget differently. An Aside session was observed opening with grep -rn although the block already banned it, so the block now says what to do when the shell is used anyway - rg, and rg cannot report completeness - and pays for it by moving two refinements into the on-demand catalog, where they already belonged: the glob-precedence sentence and the rule that a skipped file symlink does not lower completeness. Measured after that exchange: 2,019 bytes.

What that number does and does not mean. It is a byte measurement of the description, not a measurement of agent behaviour. The relocation was checked three ways: the suite pins the call-shape tokens that had to survive, a byte budget now fails the suite above 2,048 bytes, and actions.describe('search.content') was run live to confirm the completeness envelope and scope.skippedSymlinks are reachable on demand. The paired A/B trial that would show a slimmed description is not WORSE for an agent was not run, so no usability claim is made here.

The ERG404 above was recorded when the resolver had no candidate for the ripgrep that Aside itself bundles, so the fix at the time was an absolute rgPath pin. 0.7.0 added that bundled binary to the ladder, ahead of PATH and behind any explicit path, derived from the home directory on macOS and Windows.

Re-measured on macOS with no pin at all (rgPath: null): one aside exec session called mcp__aside-codemode__execute_code exactly once, ran a fixed-string search.content, and got its row back with complete:true, truncated:false and partial:[]. The agent used no other tool, so no fallback produced the answer. The explicit pin is now an override, not a requirement.

Which binary served that search is an inference, not a direct observation: an MCP response carries rows and completeness, not the resolved path. It rests on four premises checked at the same moment. The configured rgPath was null; the daemon-spawned child carries the same six-variable environment that produced the ERG404 above; and the ladder's three fixed macOS locations — /opt/homebrew/bin/rg, /usr/local/bin/rg and /home/linuxbrew/.linuxbrew/bin/rg — were all absent. What remains is the Aside bundle. Install a ripgrep into one of those locations and the inference no longer holds.

• No automatic activation: a server registration with 0 tools cached did not attach. The GUI Add flow or per-server Refresh tools action must populate mcp.inventories.
• Hot-attach behavior for an already running session is unverified. Start a new session after refreshing the inventory.
• The 6,775 -> 1,793 byte reduction is a size measurement only. No trial established that an agent performs as well with the shorter description.
• Pinless resolution was measured on macOS only. The Windows daemon was not re-measured after 0.7.0; the earlier Windows run succeeded through the package-vendored relative path instead.
• ERG404 is not structurally impossible. An explicit path that fails still throws it by design, Linux has no bundled candidate, and a moved or unexecutable bundle exhausts the ladder.
• The bundled binary serving a search is an inference from the premises above, not a reading of the resolved path out of the MCP response.

0.9.0 re-measures the same four numbers with the same method, because the completeness work moved all of them. Eleven action signatures were corrected to declare the fields their calls actually return, browse.readText gained complete, contentShape and lostTo in both of its discovery definitions, and the resident description gained the sentence naming the 64 KiB return-path cap. Measured with Buffer.byteLength over TOOL_DEF and over actions.describe() for each of the 34 rows:

• resident description 2,034 bytes, still under the 2,048-byte budget the suite enforces
• full tool definition 2,433 bytes (2,034 description, 274 input schema)
• the 34 action descriptions 44,330 bytes together, 18.2x the resident definition
• browse.exec alone 8,249 bytes

What it does not mean, same as before: these are byte measurements of text, not measurements of agent behaviour. The ratio is the argument for fetching action descriptions on demand rather than shipping them resident, and nothing more.