Hordecode is a terminal-native AI coding agent. It is a friendly fork of opencode that tracks upstream, carries a delta-port of MiMo Code, and adds its own layer: a supervisor for long-running work, a harness pipeline, an inbox, persistent memory with an optional on-device embedding index, skills, workflows and cron.

Bring your own API key: Hordecode works with the more than 200 providers in its model catalog and with local models. Every opencode provider, plugin and config file keeps working. Hordecode is maintained by Silicon Stencil and is not affiliated with or endorsed by either upstream project.

Why Hordecode

The supervisor, the harness and research pipelines and memory run inside the session runtime, so they share the session's permissions, sidebar and storage instead of living in a separate plugin process.

• Long-running work is built in. A supervisor runs delegated work as tracked child sessions with waiters and task links, so multi-step work survives past one turn. Tasks, workflows, cron and an inbox are tools the agent calls, not plugins to assemble. Shell commands in scheduled jobs run only after you approve them.
• Memory that stays local. Seven memory tools keep project knowledge across sessions. Keyword search (BM25) works out of the box; semantic search is opt-in and runs on-device.
• Skill pipelines ship with the binary. A harness set of 16 skills (plan, tdd, debug, review, verify, ...) and a separate research set of 16 skills.
• Visible, not a black box. Running agents, subagents and scheduled jobs show in the sidebar with status and elapsed time.
• A maintained base. The core, CLI, server and TUI track the opencode 1.x line, so providers, plugins, MCP servers and existing config keep working.

Base and upstream tracking

| Part | Follows | Notes | | --- | --- | --- | | core, CLI, server, TUI, SDK and the small packages | opencode 1.19.3 | ported as a whole at each base bump | | web, desktop and shared UI packages | opencode 1.17.18 | ported in their own step | | supervisor, harness, memory, inbox, cron, workflow, skill, task and the other fork packages | Hordecode | fork-only; supervisor and harness are customized ports of the MiMo Code designs (parity 0.1.11) |

Hordecode 0.1.x stays on the opencode 1.x line; each base bump ports a newer 1.x release whole. The file-level ownership map is in docs/UPSTREAM-MAP.md: 1,161 files are Hordecode's own and 691 upstream files carry Hordecode changes. How Hordecode follows upstream releases is in ROADMAP.md.

Quick Start

# npm (all platforms)
npm install -g hordecode

# One-line install (macOS / Linux)
curl -fsSL https://hordecode.com/install | bash

# Run
horde

hordecode is an alias of horde; both commands are installed. On first launch, add a provider with horde providers (or horde auth), then start a session.

Upgrade an existing install with the method it was installed with:

horde upgrade            # latest
horde upgrade 0.1.1      # a specific version

[!WARNING] 0.1.x is a pre-release. APIs, config and storage may change without a migration path.

https://hordecode.com/install redirects to the install script in this repository; the raw URL below works the same for mirrors and pinned versions. The script checks the release archive against the release's SHA256SUMS, installs into ~/.hordecode/bin and adds that directory to your shell profile unless you pass --no-modify-path.

curl -fsSL https://raw.githubusercontent.com/siliconstencil/hordecode/main/install | bash -s -- --version 0.1.1
curl -fsSL https://raw.githubusercontent.com/siliconstencil/hordecode/main/install | bash -s -- --no-modify-path
./install --binary /path/to/hordecode     # install a locally built binary

Set HORDECODE_RELEASE_BASE to install from a mirror of the GitHub releases.

| Target | npm package | Notes | | --- | --- | --- | | linux-x64 | hordecode-linux-x64 | verified target; ships the on-device embedding index | | linux-x64 (no AVX2) | hordecode-linux-x64-baseline | | | linux-x64 musl | hordecode-linux-x64-musl, -baseline-musl | Alpine and other musl systems | | linux-arm64 | hordecode-linux-arm64, -musl | | | darwin-arm64 | hordecode-darwin-arm64 | | | darwin-x64 | hordecode-darwin-x64, -baseline | keyword memory search (BM25) | | windows-x64 | hordecode-windows-x64, -baseline | install with npm | | windows-arm64 | hordecode-windows-arm64 | use npm |

The hordecode npm package selects the right platform package at install time. Every target has keyword memory search (BM25); linux-x64 adds the on-device embedding index.

What Hordecode adds

Fork features surface as tools the agent can call, not as extra CLI commands:

• supervisor: runs delegated work as supervised child sessions with spawn references, task links and waiters, so multi-step work is tracked past one turn.
• task and workflow: units of work, and declared multi-step workflows.
• memory-add / memory-recall / memory-update / memory-session / memory-status / memory-dump / memory-gc: persistent project memory. With memory.embedding enabled (linux-x64) recall uses an on-device embedding index; otherwise BM25.
• skill and skill-search: discover and load skills from the user and project skill directories; the bundled research pipeline ships as a skill set.
• cron: scheduled runs.
• session: manage sessions and worktrees, including a fleet view of running agents.

Behind the tools, the inbox carries messages between agents and sessions, and the harness pipeline drives supervised runs.

The upstream tool set (read, edit, bash, grep, glob, lsp, MCP, plan, question, ...) and the built-in build, plan and general agents are unchanged. Hordecode adds two primary agents, harness and research. Switch agents with Tab.

Harness and research pipelines

Pick build for a direct change, harness for build work that should go through a fixed discipline, and research for a question that needs sources and a decision.

• harness loads the 16 harness:* skills: ask, brainstorm, debug, deliberate, eval, execute, feedback, merge, parallel, plan, report, review, subagent, tdd, verify, worktree.
• research loads the 16 research:* skills and runs ask, scope, gather, verify, filter, synthesize, decide and eval. Gathering runs up to five sub-agents in parallel, each with its own context; the report is written once from the verified evidence, with a citation table, and every run ends with a decision record.

Memory search: keyword by default, semantic opt-in

Memory recall is keyword search (BM25) out of the box and needs no download. Semantic (vector) search is opt-in. Enable it in hordecode.json:

{ "memory": { "embedding": true } }

On first use Hordecode downloads the embedding model once (about 133 MB) into its cache directory; recall uses keyword search while the model loads, then becomes hybrid (keyword + semantic). The on-device index ships in the linux-x64 package; other targets use keyword search. memory-status reports the current mode and vector count, and horde debug embedding-check verifies the pipeline.

Network access

Every outbound connection, what starts it and how to turn it off is listed in docs/NETWORK.md. Hordecode has no product telemetry.