2026-05-30

Hermes Ships Tool Search, MCP Optimization, and Infrastructure Hardening in One Wave

hermestool-searchmcptoken-optimizationatomic-writes

Hermes Agent shipped a series of updates on May 30 that combine a headline token optimization with infrastructure hardening. The changes span tool schema loading, MCP server startup behavior, file system writes, kanban resilience, and skill synchronization ordering - all covered by @InfomlyLab in a thread of posts.

Hermes Agent killed the per-turn token tax that ships with every MCP tool.

The old design dumped every tool schema into the prompt on every turn. Fifty tools. Twenty thousand tokens. Before a single line of work.

Three bridge tools replaced the entire schema dump: tool_search,…
— Infomly (@InfomlyLab) May 30, 2026

Tool Search: 41% → 3% Context Overhead

The previous design loaded every tool's JSON schema into every turn's system prompt. With 50+ tools - browser actions, file operations, cron management, Discord integration - a single turn started with ~20,000 tokens of tool definitions before any work began. For a 50-tool setup, tool schemas consumed 41% of the available context window.

Tool Search replaces the schema dump with three bridge tools:

tool_search - queries available tools by name or description
tool_select - loads a specific tool's full schema on demand
tool_clear - unloads schemas when no longer needed

Core tools (terminal, memory, file operations) remain always-loaded. Everything else is fetched when the agent determines it needs them. In a 16-tool scenario, context overhead dropped to 3%.

The implementation works because the agent already knows its task domain from the user's request. A file-system operation triggers a tool_search("write_file"), which returns the schema, which the agent uses once and either keeps or clears. The model never sees schemas for tools it does not use.

Hermes killed the 7.5-second MCP freeze at TUI startup.

One slow MCP server held up the whole agent. No loading bar. No fallback. Just dead air.

Now the agent launches in 115ms. MCP discovery runs in the background. Fast servers land in time. Slow ones don't block anything.

If…
— Infomly (@InfomlyLab) May 30, 2026

MCP Startup: 7.5s → 115ms

MCP servers could block agent startup entirely. A single unreachable or slow MCP server consumed its full 7.5-second retry window before the TUI appeared - no loading indicator, no fallback, no progress.

The fix runs MCP discovery asynchronously in the background. The agent launches immediately at 115ms regardless of MCP server state. Fast servers connect during startup; slow or dead servers arrive when they arrive, and they do not block anything. The status bar updates incrementally as servers connect.

This changes the failure mode: a dead MCP server now means one missing tool, not a frozen agent.

Atomic File Writes

Every file write operation in Hermes now goes through a two-phase commit: write to a hidden temporary file first, then atomically rename over the target. Same-filesystem rename is an atomic operation on every operating system - the target file transitions from old to new in a single filesystem operation with no intermediate state visible to readers.

A crash or kill signal during a write now leaves either the complete old file or the complete new file. The half-written-corrupt-file failure mode no longer exists.

Kanban Loop Fix and Skill Corruption

Two bugs received fixes in the same wave:

Kanban infinite retry. When a kanban task failed, the retry mechanism had no memory of previous attempts. It would try again - and again - indefinitely. The fix tracks failure counts and stops retrying when the limit is reached.

Silent skill corruption. The skill management pipeline had an ordering bug: it deleted the manifest database entry before touching the filesystem. When directory deletion failed on a read-only tree (common with Nix installations), the function returned early. The manifest entry was already gone, but the user's copy remained intact. Future sync operations skipped the skill because the manifest said it did not exist. The fix reorders: filesystem first, manifest update second.

Context compression also received throttling logic to prevent performance thrashing in long tool-heavy conversations - when rough estimates were noisy but real usage fit within the window, the agent now skips redundant compression cycles.

Community: Local-Only Hermes

Imagine Hermes Agent using Actual Computer running Qwen-3.6-27B with the fastest MTP support. If you cool down an M5 Max enough you can get 40 tok/s out of it no problem. Internet goes down, agent stays on. Miraculous to a freak like me
— ∿ (@somewheresy) May 30, 2026

@somewheresy posted about running Hermes fully locally on a MacBook M5 Max with Qwen-3.6-27B and MTP support at 40 tokens per second. The setup survives internet outages - the entire stack runs on-device with no external API dependency.

On the UI side, @Krongggggg shared a local web dashboard for Hermes that visualizes kanban task states, the skill catalog, cron jobs, and memory files - pulling directly from the ~/.hermes configuration directory with no additional setup.

The data across these updates points to a single pattern: infrastructure hardening shipped alongside the headline feature, not after it. Tool Search reduces token cost; atomic writes, MCP async discovery, and kanban/skill fixes reduce failure modes. A wave of updates where the secondary changes are as structural as the primary one.

[^1]: Infomly. "Hermes Agent killed the per-turn token tax." X. May 30, 2026. [^2]: Infomly. "Hermes killed the 7.5-second MCP freeze at TUI startup." X. May 30, 2026. [^3]: Infomly. "Hermes Agent made file writes atomic." X. May 30, 2026. [^4]: Infomly. "Hermes kanban had a bug that let broken tasks run forever." X. May 30, 2026. [^5]: Infomly. "Hermes Agent fixed a silent skill-corruption bug." X. May 30, 2026. [^6]: ∿. "Imagine Hermes Agent using Actual Computer running Qwen-3.6-27B." X. May 30, 2026.