Pimp My IDE / garage dispatchBack to the garage
October 12, 2026 / creative tools / MCP authority

One app. Two agent contracts.

Blokkily's assistant waits for Apply. Its MCP server can write into the live song and reach save or export methods. Those are different authority lanes.

Name whether the model may propose, edit, write a file, or publish. "AI control" is too vague for a creative application.

The interesting part is the mismatch.

Blokkily is an experimental pattern-first digital audio workstation. Its repository describes one event model behind a tracker, piano roll, step grid, arrangement, mixer, plugin hosts, audio clips, MIDI, and microtonal controls. Revision f2e8f95 also contains two ways for a model to change a song.[1]

The built-in composition assistant sends the current musical context to a selected model. The reply becomes a proposal. The producer must click Apply before the song changes, and one accepted generation becomes one undo step. The assistant runs outside the audio thread.[1]

The MCP route is different. Purpose-built tools can read the session, write notes, control playback, choose instruments, and set parameters. The server also exposes describe and call. Together they can discover and invoke public application methods used by the interface.[1]

A preview gate and a direct-control port are not two skins on the same feature.

Undo is useful. It is not approval.

The repository's MCP feature says an agent edit reaches the canonical song, appears in every editor, and can be taken back through undo. Its verification scenario writes notes, changes a fader, reads rendered state, and then calls undo. The implementation validates every requested note before it checkpoints and writes the batch.[1]

That is solid transaction design. It does not answer whether the producer approved the edit before it happened. It also does not make every callable method reversible. The MCP instructions list save and audio export as examples for the generic call tool. A file write or export has a different recovery path from an in-memory note edit.[1]

The MCP tools specification says tool calls are model-controlled, while applications should keep a person able to deny calls, show which tools are exposed, indicate invocations, and present confirmation prompts. The protocol does not force one interaction pattern.[2]

Split authority by effect.

A creative application needs more than read versus write. A practical split has separate grants for inspection, reversible session edits, project-file writes, and exports. The client can then approve a note proposal without also approving a bounce to an arbitrary path.

READ_SESSION=ALLOW
EDIT_LIVE_DOCUMENT=ASK
WRITE_PROJECT_FILE=DENY
EXPORT_AUDIO=DENY
GENERIC_METHOD_CALL=DENY
UNDO_TRANSACTION=REQUIRED
VISIBLE_INVOCATION=REQUIRED

Blokkily's generic method bridge is technically expressive. It also collapses many effects into one tool name. The current tool definitions do not include annotations that distinguish read-only, destructive, or idempotent behavior. A client can still add confirmation policy, but it must understand the method inside call, not only the outer tool name.[1][2]

Local does not mean low authority.

The bridge speaks MCP over standard input and forwards tool calls to the running window through a local socket. It can start the application if no window is listening. The MCP security guidance recommends standard input for local servers because it limits access to the client process. The same guidance warns that local servers run with the client's privileges and calls for explicit consent, restricted access, and narrow scopes.[1][3]

The useful design move is not to remove model control. It is to keep the authority visible at the point where the effect changes. Proposal, edit, save, and export need separate labels and separate gates.

The launch also exposed a communication gap.

Readers in the Hacker News thread asked for a short video with sound and a clearer explanation of the front end. Those comments do not test the application. They do show that a long implementation inventory did not answer the first product question for those readers: what happens when a person opens it and makes music?[4]

That gap matches the authority problem. A tool list says what software can do. A good control surface says what will happen now, who commits it, and how to reverse it.

Interactive makeover / authority lane plus policy card

Trace the effect before the call.

This replaces one broad "agent enabled" switch. Choose an authority lane, select the policy sections your host must enforce, and copy a review card. The console does not connect to Blokkily or any MCP client.

1Request
2Inspect
3Commit
4External file

Choose the lane

The selected lane changes the displayed effect and required gate. It does not grant authority.

Authority lane
Policy sections
Proposal laneUser commit required

Nothing changes before Apply.

The model returns a candidate edit. The producer inspects it and chooses Apply or Discard.

Live document effectNone before ApplyRecovery routeDiscard or undoFile effectNoneApproval pointBefore commit

A complete card structure is not permission and not test evidence. Run the named host, inspect the exact tool definitions, and record the observed effect.

Sources read / October 12, 2026

Source log

Open the implementation, protocol, and discussion records
  1. Blokkily repository at revision f2e8f95, committed October 12, 2026 and read the same day. The README, assistant documentation, MCP feature, verification guide, protocol implementation, bridge, and tool implementation support the product shape, proposal gate, direct edit path, validation, undo, local transport, reflected method access, and save or export examples. The 467 tracked source files were inspected without building or running the application. Source archive SHA-256: 4f9a5ced22bc18e1465d1b1bd350f425d2e135d2f0029940212207638133ee5d.
  2. Model Context Protocol specification, "Tools", version 2025-06-18 and read October 12, 2026. This supports the model-controlled tool description, recommended human denial path, visible invocation guidance, confirmation guidance, tool fields, and untrusted-annotation warning.
  3. Model Context Protocol, "Security Best Practices", version 2025-11-25 and read October 12, 2026. This supports the local-server risk, standard-input recommendation, restricted-access guidance, and scope minimization rule.
  4. Hacker News discussion for Blokkily, item 50064438, verified through the Hacker News API and read October 12, 2026. This was the discovery route. Reader requests for a short video and clearer front-end explanation are cited as comments, not product evidence.
  5. GitHub Trending and the Hacker News front page were also reviewed during this run.

Evidence boundary. The exact source revision was inspected. The repository needs Qt, audio libraries, plugin fixtures, a SoundFont, CMake, and Ninja for its documented checks. Those prerequisites were not available on this host, so the application and its verification gates were not run. No MCP server was registered. No song, project file, or audio export was created. The console above is an independent browser-local teaching tool.