HYTHE
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@HYTHEresume from the last checkpoint and report the current task status"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
HYTHE
A coordination bus + shared truth for your agent fleet. Self-hosted, two-component setup: a dockerized server plus a thin stdio bridge for your MCP clients.
HYTHE is an MCP server that lets multiple AI coding agents — Claude Code, Codex, Cursor, custom harnesses — share state, preserve context across sessions, and coordinate with each other. It is not another memory store. Its differentiators:
Supersession + current-state resolution: observations supersede each other (
replace-current), andget_current_observationresolves the chain server-side — so readers get the newest non-superseded state. Conflict handling is implemented:checkpointis branch-preserving CAS — concurrent writers branch rather than overwrite, and conflicts surface as heads, never silently resolved (contract-tested; see the evidence ledger).Agent messaging with a tracked delivery lifecycle: direct and capability-based sends, message supersession, read-state as a shared signal, atomic ack+archive. (Described as tracked lifecycle, not guaranteed delivery — the states are honest about what the server knows.)
The Adaptive Coordination Protocol (ACP): the ETA-driven coordination discipline that co-evolved with the server — published as docs/SPEC.md with a real two-harness worked tutorial.
Documentation
Doc | What it covers |
Clean machine → two coordinating agents in ~15 minutes | |
One current truth, messaging lifecycle, resume/checkpoint, honest security model | |
Current-head pointer, reconciliation, record / patch, the v3 read model, and the one-time adoption step for scopes written before v3 | |
The coordination protocol, versioned (1.0.0) | |
Annotated transcript of a real Claude Code ↔ Codex review loop from the original Engram deployment | |
The 20-tool v1 surface; every retired tool and its exact replacement | |
Tested SQLite backup/restore/compaction runbook | |
Offline issuance, attestation, promotion, rotation, and revocation | |
Hash-bound, owner-approved disposition of ambiguous historical rows | |
Verified offline VACUUM and no-clobber promotion after logical cleanup | |
Pinned deployment, readiness, shutdown, canary, and rollback contract |
Related MCP server: samskriti-project
Status
Version 0.2.0. This source tree and package define the reviewed 0.2.0
release. Version 0.2.0 preserves 0.1.7's dual-proof agent authorization,
exact-mailbox/private-payload containment, bounded related-context discovery,
latency-SLO recovery and the offline adjudicator, and adds checkpoint /
resume resultVersion: 3: an explicit current-head pointer per scope, a
reconcile operation that folds live heads into one survivor and resolves
divergent values causally, append-only verified fact/loop versions, a read
model that only promotes values proven on the accepted lineage, and
record / patch operations on the pointed head. Every v1 / v2 request and
result shape is frozen; v3 is opt-in per call, so a 0.1.7 client keeps working
against a 0.2.0 server. The dependency tree moved to Express 5, clearing two
moderate qs advisories. This release runs on the reference production host
since 2026-09-05. All 0.2.0 rollout configurations pin @hythe/mcp@0.2.0;
verify that the registry returns that exact version before production use. The
predecessor @tomcat65/engram-mcp and io.github.tomcat65/engram remain
available as compatibility history and are never unpublished. See
CHANGELOG.md and the v3 guide.
House rule: every claim in these docs must trace to a test or a measurement (see the evidence ledger below). Claims that don't are bugs.
Evidence ledger
Claim | Evidence | Source | Date |
Tree green, full gates | Source and clean installed-package gates cover 49 Vitest files (530 passed, 2 skipped, 6 explicit todos), 107 migration checks, 43 release-tree checks, 31 agent-kit checks, zero production npm audit findings, registry validation, and an isolated Docker start/readiness/MCP/WebSocket/graceful-shutdown canary. |
| 2026-08-14 |
Compaction identity fails closed | Startup and post-compaction hooks reject missing, conflicting, or invalid identity; the bridge binds acting tools and message-resource recipients to the configured exact lane before any HTTP request. |
| 2026-08-13 |
Exact mailboxes stay isolated | Houston/Hythe, case variants, display metadata, legacy identity history, cross-tenant poisoning, suffix handles, lifecycle changes, supersession, HTTP, and ENG-4 authorship have negative isolation coverage. |
| 2026-08-13 |
Private message bodies stay out of shared graph surfaces | Historical, oversized, malformed, alias-linked, relation-linked, imported, exported, searched, graphed, logged, and restored payload paths are covered; hidden children do not leak through counts. |
| 2026-08-13 |
Cleanup is offline and fail-closed | Migration 007, owner-approved adjudication, and physical sanitation are dry-run/plan-first, bind execution to reviewed database and decision evidence, require verified no-clobber artifacts, refuse unresolved custody, and preserve rollback state. |
| 2026-08-14 |
Per-agent authority is server-derived | The base key proves deployment/tenant access; a separate protected-file |
| 2026-08-14 |
Install and discovery surfaces agree | CLI and registry surfaces require an exact lane identity and expose only a protected per-agent key-file path plus an explicit authorization mode; tools and resource templates served through HTTP match the stdio bridge path. |
| 2026-08-14 |
Current head is explicit, never guessed | The pointer advances only from the pointed head; stale parents branch without becoming current; a pointer at a non-live snapshot fails closed; snapshot rows are immutable by trigger. 33 contract tests. |
| 2026-09-03 |
Versions and coverage are exact and verified | Every ledger tuple gets exactly one disposition; the startup backfill is all-or-nothing and idempotent; a bidirectional parity verifier fails closed on any missing, extra or mismatched row. 32 contract tests. |
| 2026-09-03 |
Reconciliation is CAS-bound and causal | Exact live-head set and pointer are compare-and-set; every non-survivor head is retired (never deleted); divergent values resolve only by bound accept, reject or lineage reject; the record is payload-bound and re-verified on replay and read. 34 contract tests. |
| 2026-09-04 |
The v3 read model promotes only proven values |
|
| 2026-09-04 |
record / patch cannot promote stale state | Only the pointed head is admitted (conflict otherwise); state is materialized from the hash-verified payload; each payload is read once per call; snapshot resources serve exactly the verified bytes. 21 contract tests, four review rounds. |
| 2026-09-04 |
Express 5 keeps the HTTP contract | A bodiless or wrong-content-type request is handled like an empty JSON body on every parsed route, never a framework 500; production audit reports zero findings. |
| 2026-09-04 |
Production migration verified on a copy first | The v3 schema and backfill were exercised against a copy of the production database on the production host before cutover; the live cutover took 17 s with row counts identical to the verified cold backup. |
| 2026-09-05 |
v1 surface (short)
Install (two-component setup): server via
docker compose up, then the@hythe/mcpstdio bridge per MCP client; first-run wizard stores the API key in a protected.envand prints secret-free file-referencing config for Claude Code / Codex / Cursor / Claude Desktop; optional namespaced (demo-*) demo seed on an otherwise empty DB.Knowledge graph: entities, observations, relations, supersession, current-state resolution, conflict surfacing.
Related-context discovery: exact scope anchoring plus automatic bounded vector retrieval and graph-path reranking, with currentness/evidence/provenance explanations and no implicit writes.
Agent messaging: direct, capability-based, superseding; tracked delivery lifecycle.
resume/checkpoint: budgeted one-call session rehydration with closed coverage accounting, and CAS-protected, branch-preserving structured state capture. WithresultVersion: 3: an explicit current-head pointer,reconcileto fold heads and resolve divergent values, verified fact/loop versions with provenance, andrecord/patchon the pointed head — see the v3 guide.ACP: the coordination protocol as a versioned spec + real worked example.
Embedding runtime contract
The published npm package does not install a transformer engine by default.
Without one, HYTHE still provides deterministic 384-dimension token-hash
vectors plus lexical ranking; it does not download a model or create a model
cache. @xenova/transformers@2.17.2 is an optional peer for operators who
deliberately provide it from their own audited dependency root.
The Docker image keeps semantic behavior stable through a separate, locked
/opt/hythe-transformers runtime. Its startup preflight loads the configured
Xenova/all-MiniLM-L6-v2 q8 model and requires one finite 384-dimension
embedding before the server starts. When
SQLITE_VEC_ALLOW_REMOTE_MODELS=false, the configured
SQLITE_VEC_CACHE_DIR must already contain that model. Do not switch between
transformer and hash embeddings on a populated vector index: retain the same
provider/model/dimensions or rebuild the vector index first.
License
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections. Dates show when Glama detected each change.
No tool schema history has been recorded yet.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Connectors
Cloud-hosted MCP server for durable AI memory
shared AI-context layer for teams — persistent memory your agents search and update over MCP
An MCP memory server. One memory your agents share — across models, devices and apps.
One memory, every AI. A shared, user-owned markdown memory your AI clients read and write over MCP.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceA shared memory and coordination server for multiple AI coding agents, built on the Model Context Protocol (MCP).5MIT
- AlicenseAqualityBmaintenanceA local MCP server that enables multiple AI coding tools to share structured project state (decisions, tasks, bugs) so they coordinate without re-explaining.5MIT
- AlicenseNot gradedqualityDmaintenanceA self-hosted MCP server enabling multiple AI coding agents to share state, preserve context across sessions, and coordinate with each other.40Apache 2.0
- FlicenseNot gradedqualityBmaintenanceCentralized multi-project state & handoff MCP server that lets AI agents save checkpoints in one session and resume in another with minimal token overhead, including git snapshots and code-graph integration.-
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/hythe-dev/hythe'
If you have feedback or need assistance with the MCP directory API, please join our Discord server