Terminal and files
Compresses repetitive stdout, bounded file reads, and safe Get-Content calls without hiding failures or exact recovery.
Codex plugin · MCP server · local-first
Capsule is a Codex token efficiency tool and MCP context compression layer. It shows the model the smallest safe evidence view, while keeping the complete local result available on demand.
Compresses repetitive stdout, bounded file reads, and safe Get-Content calls without hiding failures or exact recovery.
Compiles symbols, dependency impact, durable facts, and context handoffs before the model rereads an entire codebase.
Keeps structured results, URLs, and references recoverable while limiting repeated catalog and page payloads.
Token-negative, unsafe, failed, literal, or explicitly full requests pass through unchanged.
Every result follows a simple local pipeline:
request → classify → compact evidence view
↘ exact local capsule → expand on demand
Capsule's published A/B figures are workload-specific model-visible measurements, not a promise about hidden reasoning, provider caching, billing, or subscription quotas.
Read the implementation, tests, release history, privacy contract, and reproducible benchmarks in the public repository. Contributions and security reports are welcome.
Browse all Capsule guides from one crawlable index.
Understand what Capsule can reduce, what it cannot observe, and how to measure it locally.
See how the single Capsule surface handles terminal, files, projects, web, memory, and recovery.
Read the bounded replay methodology, safety gates, and exact-recovery results.
Follow the portable Codex plugin and MCP setup, verification, and troubleshooting path.
Verify the bundled map-token-context skill, plugin manifest, and no-match routing safety.
Use canonical links, release-aware references, and measured claims when sharing Capsule.
Pair real workloads, compare model-visible payloads, and report recovery and fallback limits.
Measure metadata, traffic, referrals, releases, and repository-search snapshots without ranking claims.
Keep web results compact, exact-recoverable, and measurable across first and repeated calls.
Project repeated stdout/stderr safely while preserving failures, structured evidence, and recovery.
Get concise answers about observable savings, exact recovery, compatibility, privacy, and measurement boundaries.