Skip to content
PidocsCLI & tooling

Compare Pi's native read tool to Amp's tool/MCP model

Session from pine-of-glass-sessions on Hugging Face by thomasmustier · MIT · condensed by Coders Talk

6msession
0interventions
GOAL

Understand what Pi's built-in read tool does versus letting the model use bash directly, then compare against how Amp (a competing coding agent) handles images, tool discovery, and MCP.

  1. Prompt · +0
    What does Pi's native read tool do, and do we need it if models can just cat/head files like Amp allows?
  2. Agent did · +0m
    Read Pi's README, docs, and read.js/read.d.ts source to explain what the built-in read tool does: line slicing, truncation limits, image attachment support, extension overrides.
  3. Prompt · +2m
    What does the read tool do for images that plain bash doesn't?
  4. Agent did · +2m
    Explained that read detects image MIME types and sends them as actual model image attachments, unlike bash which only surfaces text/bytes.
  5. Prompt · +2m
    How does Amp handle image files?
  6. Agent did · +2m
    Searched web for Amp's docs and summarized that Amp treats images as user-attached context (paste or @-mention) rather than a read-tool fallback.
  7. Prompt · +3m
    Check the Amp docs post about Opus 4.8, and pull the Amp CLI code from npm/Sourcegraph to verify.
  8. Agent did · +3m
    Fetched Amp's news/manual pages, npm-packed the @ampcode/cli and darwin-arm64 binary, then ran `amp tools list/show` across modes to find that Amp dropped the text Read tool from smart mode but kept a separate view_media tool for images/PDF/audio/video.
  9. Prompt · +5m
    How does Amp do tool discovery / progressive disclosure / tool search, and what's the read_mcp_resource tool about?
  10. Agent did · +5m
    Searched Amp's blog posts and a third-party writeup, then ran `amp tools show` on read_mcp_resource, skill, and read_web_page to conclude Amp replaced general tool search with skills as the discovery layer plus lazy-loaded MCP tools.
  11. Outcome · +6m
    Session ended with a full comparison of Pi's read tool vs Amp's image handling and skills-based tool discovery, backed by inspected source and docs, no code changes made.
VERDICT

I'd start next time by directly asking to pull and diff the actual CLI source for both agents instead of starting with docs, since that's what actually settled the questions (unpacking the npm tarballs and running `amp tools show`). The npm-pack-and-inspect approach worked well for getting ground truth beyond marketing docs, and no real interventions were needed since this was pure research with no wrong turns to correct.

I tried this · 0

Discussion · 0

How Builds work