Skip to content

Latest commit

 

History

History
65 lines (49 loc) · 3.92 KB

File metadata and controls

65 lines (49 loc) · 3.92 KB

Token-saver landscape — 2026-07-13

Public GitHub metadata snapshot. Stars are discovery signals, not benchmark scores.

Ranked candidates

Tool Stars Freshness Evidence verdict
Graphify 83,830 pushed 2026-07-13 Excellent persistent graph projection; not an always-on saver
Ponytail 81,839 v4.8.4, 2026-07-10 Popular; local A/B was net more tokens
RTK 70,717 v0.43.0, 2026-06-28 Strongest mature default; stable process-fixture reduction 97.25%
CodeGraph 59,586 pushed 2026-07-13 Strong repeated repo relationship queries; workload-gated
Headroom 58,872 v0.31.0, 2026-07-09 Large vendor claims; observed Codex gain was workload-dependent
context-mode 18,878 v1.0.169 99.91% large-log projection; 7,458-token cold schema
LLMLingua 6,431 pushed 2026-04-08 Mature research compressor, not a ready coding-agent stack
lean-ctx 3,230 v3.9.8, 2026-07-12 Interesting but 76 MCP tools and <14-day release
Tilth 310 v0.9.0 Local code-read reduction 86.43%
squeez 167 v1.36.0, 2026-07-10 Unverified 95% claim; <14-day release
token-saver 115 pushed 2026-06-02 Useful format filters; no measured win over RTK/native projection
mcp-compressor 97 v0.31.5, 2026-07-09 Good schema-router design; unnecessary for one lean MCP
tokbench 9 pushed 2026-06-15 Useful eval direction; low adoption

Community sources:

The strongest repeated community pattern is progressive tool/schema disclosure. Manufacturer savings claims were not promoted over local accepted-output measurements.

Local versions and upgrade decision

Tool Local Upstream Decision
RTK 0.43.0 0.43.0 current
context-mode 1.0.169 runtime; older Codex cache 1.0.169 runtime current; keep on demand
Headroom 0.31.0 0.31.0 optional provider/proxy; not Lean-default and never MCP
Tilth 0.9.0 0.9.0 release current release
Graphify 0.9.11 0.9.14 defer: release inside 14-day gate
CodeGraph 1.3.1 1.4.1 defer: release inside 14-day gate
Ponytail local unversioned skill 4.8.4 do not global-load; measured net negative

No new candidate was installed merely because it was new or popular.

Graphify query measurement

Existing code graph:

  • 2,982 nodes, 6,187 links, 10 hyperedges.
  • raw graph.json: 3,094,806 bytes, about 773,702 tokens.
  • graphify query --budget 800: about 652 tokens, 450 ms.
  • projection: 99.92% smaller, 1,186.7x reduction.

The graph was stale versus the repo and used a pre-#1504 node-ID schema. This proves compression, not current answer accuracy; rebuild after the Graphify upgrade passes the age gate.

Decision

default: router -> RTK -> native projection -> memory/Tilth CLI on demand
heavy payload: context-mode session
repeated structure: choose Graphify OR CodeGraph per repo
broad MCP catalog: benchmark mcp-compressor before enabling

Do not combine every layer. The measured all-on profile cost 12,567 input tokens versus 1,918 for CLI-selective.