OpenCode++
A verification control plane around OpenCode Desktop that makes AI diffs, boundaries and evidence reviewable.
- Source repo
- whut09/opencode-plusplus
- Stars
- ★ 113
- Last updated
- 1mo ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 52/100 · Major gaps
At a glance
- How it runs
- Works with
- Platform-specific
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A Windows developer using OpenCode Desktop who wants every AI change backed by auditable evidence.
- Not a fit if
- Teams not using OpenCode Desktop
- Anyone needing OS-level sandbox isolation
- Users working on macOS or Linux desktops
- Source review
- 52/100 · Major gaps
What does this agent do, and when should you use it?
OpenCode++ is a Windows-first Harness plugin for the official OpenCode Desktop application, shipped as a per-user EXE installer that needs no Administrator rights and does not patch app.asar. Installing it adds one primary mode called OpenCode++ to the mode picker; the currently selected OpenCode model then calls the in-process plugin tools instead of spawning a second model or a CLI process. Around the model's own file reads, edits and command runs, it supplies deterministic context retrieval, edit boundaries, command and protected-path guards, execution evidence bound to the current working-tree hash, and policy, freshness, regression, hallucination and convergence gates. Runtime artifacts land in the repository's .agent-context/ directory (traces, runs, loops, sidecar summaries). The plugin works offline by default, fetches no remote context and exposes no hidden chain-of-thought.
The plugin drives a task through Plan → Prepare → Retrieve → Execute → Collect → Evaluate → Decide → Persist → Finalize. prepare sets task boundaries and required checks; retrieve selects relevant files, symbols, versions and dependencies before blind search and records rejected files with reasons; guards check commands and protected paths; command or CI results are then matched to the current working-tree hash and the active evidence policy. evaluate and next return the next allowed action — repair, repack, human review or finalize — with a compact ✓ Verified / ✗ Repair required / ⚠ Human review status. The opencode_plusplus_dashboard tool exposes stage progress, selected and rejected files, decision basis, evidence freshness, intervention counts and the final summary. An Intervention Ledger tracks observed, prevented, requested, repaired, verified, unresolved and human-review states.
- A Windows developer using OpenCode Desktop who wants every AI change backed by auditable evidence.
- A team whose agent keeps declaring success without running tests, and who needs a fresh-evidence gate.
- A maintainer who wants to confine edits to specific directories and block dangerous shell commands.
- An engineer who wants a reviewable trace of decisions and executions left inside the repository.
- A team that wants to customize retrieval ranking, command rules or stopping policy and redistribute their own checksummed installer.
How do you install or deploy this agent?
Download the installer from GitHub Releases:
opencode-plusplus-setup-win-x64.exeFully exit OpenCode Desktop, then double-click the EXE and accept the installation message. It installs per-user on Windows x64 and requires no Administrator permission.
Only the following OpenCode configuration files are written (default config directory %USERPROFILE%\.config\opencode; OPENCODE_CONFIG_DIR takes precedence):
<OpenCode config>\plugins\opencode-plusplus.js
<OpenCode config>\agents\opencode-plusplus.md
<OpenCode config>\opencode-plusplus\state.json
<OpenCode config>\opencode-plusplus\installation.jsonIt removes Slash Command files or app.asar patches created by older releases and no longer changes the OpenCode Desktop bundle.
How do you use this agent?
Restart OpenCode Desktop, open a repository, and select OpenCode++ in the mode picker at the bottom of the prompt box. Describe the task normally, for example:
Fix the login timeout and add a regression testLet the selected mode call prepare, retrieve, evaluate and next while it works, and do not switch back to Build for a task that needs the Harness gates.
Read the compact status returned by evaluate or next, and call the dashboard tool when you want stage progress, selected and rejected files, decision basis, evidence freshness and interventions. Inspect the repository directory for traces, findings, required commands or the final report:
.agent-context/traces/
.agent-context/runs/
.agent-context/loops/
.agent-context/sidecar/latest.md
.agent-context/sidecar/visualization.jsonContext cache and registry usage live under .agent-context/cache/ and .agent-context/context-registry/usage/; these local artifacts should normally remain uncommitted.
What are this agent's strengths and limitations?
- Draws a hard line between a suggested fix and a verified fix: only fresh command or CI evidence for the current working tree counts, not annotations, older test runs or model summaries.
- Ships as a per-user EXE that leaves app.asar untouched and writes only four config files, so install and upgrade boundaries are clear.
- Offline by default, no second model call, and every runtime artifact stays in the repository's .agent-context/ for review.
- Documents concrete extension points: the agent prompt, command and protected-path rules, retrieval ranking, evidence trust and loop stopping.
- The dashboard reports recorded system facts and decision inputs rather than dressing up model reasoning as audit evidence.
- Core operation depends on OpenCode Desktop on Windows x64; there is no documented install path for macOS or Linux desktops.
- Explicitly not an operating-system sandbox: it cannot stop other applications from editing files and cannot prove business semantics from an exit code.
- Requires OpenCode Desktop to load the configured plugin directory; antivirus locks, read-only folders and permission failures degrade to diagnostics or human review.
- CLI and MCP entry points exist but are positioned as developer/compatibility surfaces, so they are not a substitute for the Desktop dependency.
- No hosted or centrally managed policy distribution is described; team-wide policy means forking the plugin and rebuilding a checksummed installer.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| OpenCode++ This agent | 52 · Major gaps | Agent plugin / skillFree + model costs | ★ 113 | 1mo ago | TypeScript | — |
| Claudexor | 76 · Good | CLIFree + model costs | ★ 520 | today | TypeScript | Codex · Claude Code |
| Swarm | 60 · Some gaps | Agent plugin / skillFree | ★ 747 | 5mo ago | TypeScript | Claude Code |
| Calyx | 59 · Major gaps | Desktop appFree | ★ 351 | 6d ago | Swift | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Permission-boundary docs clearly separate host authorization from plugin semantic blocking with a three-state disposition, giving real least-privilege and confirmation design (2). Dependencies are declared only with caret ranges and no lockfile or audit evidence is shown, deducted to 1; source attribution is generic contributor naming in package.json/LICENSE with unverified publisher, 1. Rollback via --repair restoring only owned files without overwriting user config earns 2.
Failure reporting semantics (stale/missing/forbidden/no-progress) are explicit and well specified (2). Dependency availability rests on version ranges without lockfile evidence and the Windows EXE path depends on release artifacts unverifiable statically, 1. Self-consistency between the three-state permission model and evidence boundary is coherent (2).
Audience and scenario (OpenCode Desktop Windows users) are well scoped (2) and capability boundaries (CLI/MCP as developer surfaces, not replacing global permissions) are stated (2). Environment fit is Windows-Desktop only with no cross-platform support, 1; trigger precision relies on mode picker and prompt steering without verifiable trigger rules, 1.
Information architecture has bilingual docs with a reference layer (2), concrete install steps (2), and a full MIT license (2). No CHANGELOG and version only in package.json 0.2.12 (1); naming stability shows evolution traces via legacy allowed field alongside disposition (1); examples/FAQ are thin (1); known limitations are scattered rather than consolidated (1); maintenance responsibility points only to generic contributors (1).
Output usability is strong with compact status, Dashboard, and six audit questions (2). Marginal value and cost-benefit lack quantified benchmark results; only scripts and fixtures are visible, no result data, 1.
Claims trace to docs and code structure but lack result data or external corroboration (1); cross-source corroboration is limited to in-repo self-description (1); fact/inference separation is good, docs explicitly distinguish claims from system verification (2).
- Publisher identity is unverified by FollowAgents; verify origin and signature before running the installer EXE.
- Dependencies use caret ranges with no visible lockfile or audit evidence; assess supply-chain risk independently.
- Windows OpenCode Desktop only; other platforms or CLI-only scenarios are unsupported.
- Benchmarks provide only scripts and fixtures, no actual result data, so effectiveness claims lack quantified support.
- No CHANGELOG; version evolution and compatibility changes must be inferred from code.