7 Essential Pi Agent Components for Transparency, Cost & MCP Integration

01 pi-mcp-adapter: Access the Entire MCP Ecosystem in ~200 Tokens

Capability: Enables Pi to use any MCP server — the model discovers and invokes real tools on demand via a single proxy tool; servers start only when first called. Databases, browsers, APIs — the whole MCP ecosystem becomes available without waiting for official native support.

Highlight: Replaces "capability definition" with "proxy convergence": a single server's tool definitions can exceed 10k tokens, but this adapter registers zero per-server definitions. All calls converge into one ~200-token proxy tool; the more servers you add, the more tokens you save. Existing configurations migrate at zero cost — standard .mcp.json works out of the box; Cursor, Claude Code, and Codex configs preview via /mcp setup and import with one click.

Boundary: Tool discovery responsibility shifts to the model; if the model doesn't actively explore, the tools go unused. Server security and quality are not improved by token savings.

02 pi-lsp: Universal Language Server Socket

Capability: Connects Pi to real language servers: after every successful write/edit, the file syncs to the server and diagnostics return instantly so the model sees exact line/column errors immediately. Also registers five model tools — lsp_diagnostics, lsp_hover, lsp_definition, lsp_references, lsp_symbols — enabling type queries, go-to-definition, and reference search via the actual server instead of guessing.

Highlight: (1) Declarative configuration: one JSON declares bin, args, file match patterns, and rootMarkers; adding a language means adding config, not installing a new plugin. Each server runs one long-lived process per project root, reused repeatedly. (2) Measured feedback: diagnostics are appended, not treated as failures — writes still succeed, errors become context fed back to the model while also showing a user-visible hint.

Boundary: Language server binaries must be installed and on PATH; server startup and diagnostic wait add slight latency per write.

Best for: Multi-language projects where the model should know errors immediately instead of waiting for test runs.

03 pi-cost: Token Spend Ledger

Capability: Turns token spend into a drillable ledger: overview → project → session → message (four levels). Actual vs. estimated spend shown side by side, making it obvious where money goes.

Highlight: Implemented as a standalone local web panel (default port 5461), independent of the TUI. /cost toggles it within a session; npx pi-cost runs it detached from the agent for reviewing historical sessions without launching Pi.

Boundary: Only answers "where money went," not "how to save it."

04 pi-cache-graph: Visible Cache Hit Rates

Capability: Gives context caching a check-up: /cache graph renders a three-view TUI of hit-rate curves (per-turn, cumulative, cumulative token volume). /cache stats lists per-message token breakdown. /cache export outputs CSV for external analysis.

Highlight: Built as the observation layer for the pi-context-prune extension — want to know if pruning blows up the cache? One glance at the graph tells you. Hit-rate formula cacheRead / (input + cacheRead + cacheWrite) normalizes reporting differences between Anthropic and OpenAI.

Boundary: Observes only, does not optimize; value peaks when paired with compression extensions.

05 pi-tps: Per-Turn Speed & Cost Bill

Capability: At the end of every agent turn, automatically pops a speed bill — TPS, TTFT, total latency, input/output tokens, pause count, and blended cost rate ($/M). On session resume or /tree branch switch, auto-backfills the latest entry. /tps-export dumps telemetry as JSONL and broadcasts tps:telemetry events to other extensions.

Highlight: Bundled from the tps example extension in pi-mono; npm package has zero runtime dependencies. Timing is rigorous: TTFT measured at first real token arrival; network latency and tool-execution gaps excluded from generation speed; streaming gaps >500ms counted as inference pauses and subtracted from TPS. Three-gate speed algorithm shows — when burst delivery or unreliable timing occurs, never reporting an inflated number.

Boundary: Reports speed and rate only; root cause (model, network, queue) requires correlating TTFT and pauses. $/M depends on model pricing table; omitted when pricing data unavailable.

06 pi-inspect: Session X-Ray

Capability: Answers "what's actually in this session": loaded tools (with parameter schemas and sources), slash commands, skills, and the full initial system prompt — all laid out in a local dashboard (port 5462). Snapshots persist and are replayable.

Highlight: Share button deflate-compresses the snapshot, base64url-encodes it into the URL hash; recipient opens a static hosted page and sees it instantly — no server, no upload. Warning: link contains system prompt and cwd; don't share secrets.

Boundary: Debugging tool, not a daily driver; value concentrates in "something's off" moments.

07 pi-rtk-optimizer: Tool Output Compression Pipeline

Capability: Slims context: bash commands auto-rewritten to token-frugal rtk equivalents; verbose output from bash, read, grep compressed before feeding back to model; test, build, git, lint output each have dedicated aggregation strategies.

Highlight: Multi-stage pipeline with two safety nets: ANSI stripping → test aggregation → build error extraction → git compression → lint aggregation → grep grouping → source filtering, processed stepwise; smart truncation + hard truncation fallback. Rewrite decisions delegated to external rtk CLI; if not installed, commands pass through unchanged. Reads ≤80 lines keep byte-for-byte fidelity; outputs with hashline anchors preserve full edit anchors, never truncating the model's positioning marks.

Boundary: Depends on external rtk binary; aggressive mode strips comments and whitespace — validate on small scale first.

Project List

pi-mcp-adapter: https://github.com/nicobailon/pi-mcp-adapter pi-lsp: https://pi.dev/packages/pi-lsp pi-cost: https://github.com/NikiforovAll/pi-cost pi-cache-graph: https://github.com/championswimmer/pi-cache-graph pi-tps: https://github.com/monotykamary/pi-tps pi-inspect:

https://github.com/NikiforovAll/pi-inspect