v0.5.0 — Token compression at the gateway seam
v0.5.0 — Token compression at the gateway seam
MCPlexer now compresses the tokens it feeds into your agent's context — measured first, applied only when proven safe.
Highlights
Token compression (measure-first, zero-config)
MCPlexer compresses downstream MCP tool-result payloads before they reach the model. It auto-wires into a dry-run (shadow) mode on upgrade: it measures what each transform would save on your real traffic and returns results byte-identical — zero accuracy or latency risk — until you turn a transform on from the dashboard.
- Dashboard settings page (
/settings/compression): a three-state mode (off / dry-run / on), a per-transform toggle, and each transform's observed savings shown next to it, with a gate-verified badge for transforms proven safe. - Four transforms:
json_minify— strip insignificant JSON whitespace (value-lossless).structured_dedup— drop the double-encodedcontenttext whenstructuredContentalready carries the value.log_compact— drop low-severity INFO/DEBUG log noise, keep errors + stack traces + load-bearing tokens.oversize_truncate— head + tail for very large payloads.
- Reversible by design (CCR): any dropped content is stashed and expandable on demand via the new
mcpx__retrievetool. A verify-after-compress kill-switch guarantees the model never sees a marker it can't expand. - Measured, not asserted: the first tokenizer in the codebase + a durable savings ledger + a REST API power the dashboard. A CI gimmick gate enforces value-losslessness, credential safety, a latency budget, and a real measured win for every transform.
Also in this release
- Audit live-stream filter parity — the live tail now filters identically to the historical feed across all facets.
- macOS install — release binaries are re-signed + de-quarantined so the installed binary runs cleanly.
- Tasks — honest open-count + bulk move-to-workspace.
Good to know
- The compression seam sees MCP tool results +
execute_codeoutput; native harnessBash/Read/Grepbypass the gateway. Budget 15–35% on mixed MCP traffic. - Default is dry-run — nothing is compressed until you opt a transform in from the dashboard.
- Pure Go, no CGO, single static binary — the ML/sidecar path was deliberately left out to preserve that.
Full changes: v0.4.0...v0.5.0