Arc MCP
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Arc MCP
A smaller plan. A stronger finish. Evidence-led workflows inside OpenCode.
Arc adds 1–5 ordered outcomes, actual check records, two fresh read-only native reviewers and an advisory completion gate. OpenCode keeps its models, UI and tools.
Open source · v0.1.1
Arc MCP is publicly available under the MIT license. Explore the workflow and installation guide at https://pumkin4534.github.io/arc-mcp-site/.
Related MCP server: ape
Install
Requirements: macOS, OpenCode (tested with 1.18.25), Rust 1.88+, Node.js 20+, npm, curl and Xcode command line tools. Intel macOS is tested; Apple Silicon builds from source but has not been exercised on the development host. Linux and Windows check execution are not supported by this release.
One-line install
curl -fsSL https://pumkin4534.github.io/arc-mcp-site/install.sh | bashNo GitHub login, CLI, or repository invitation is required. The installer downloads the public versioned source, builds locally with builder paths remapped, and merges an Arc entry into existing OpenCode JSON/JSONC configuration. It supports comments and trailing commas, backs up changed configuration and refuses conflicting Arc installations. It does not replace existing providers or permissions. Source archives contain no developer sessions, logs or credentials.
Quit and restart OpenCode after installation:
/arc Implement the change with these acceptance criteria: …
/arc-status
/arc-continueInstalled binary: ~/.config/opencode/arc-mcp/bin/arc-mcp (respects
XDG_CONFIG_HOME). The installer also adds the Arc plugin, read-only agent,
workflow skill and three commands. Tests use ARC_CONFIG_HOME to select an
isolated configuration directory; normal use requires no environment overrides.
How it works
Plan: 1–5 broad dependency-ordered outcomes, retaining all acceptance details.
Build: use native OpenCode tools and your selected model.
Check: run real commands and bind bounded output/exit to scoped source fingerprints. Failed, unknown, interrupted or stale evidence blocks readiness.
Review: two fresh native sessions receive the evidence packet and read-only project access, not the writer transcript. They run concurrently.
Finish: current checks and both review judgments must agree. Progress-only updates cannot erase a failed judgment. Prior findings follow later packets.
Stop, new input, questions and permission requests interrupt reviews. Arc does not automatically resume stopped work or replay commands. One active ledger is stored per project; unfinished work cannot be silently discarded by a replacement plan.
Privacy and assurance
See PRIVACY.md and SECURITY.md.
No built-in analytics or separate telemetry service.
During use, local records include task text, paths, commands, output and reviewer IDs. Output is not automatically redacted. Keep secrets out of commands and captured output. State is not included in source packages.
Reviewer packets use your selected OpenCode model provider. A remote provider receives that task/evidence context; this is not an offline-only tool.
Completion is explicitly
native-advisory,tamper_protected:false. Ordinary native tools and same-user processes can modify host records.Only MCP-owned checks use macOS Seatbelt. Source observations are scoped and non-atomic; ignored/generated/secret content and symlink referents are excluded.
Cleanup owns a process group, not deliberately detached descendants. Run foreground tests, not daemons. Guarded checks deny network/Mach delegation and writes outside project/scratch (except
/dev/null). Install dependencies first.
Development
npm ci --ignore-scripts
cargo test
cargo clippy --all-targets -- -D warnings
npm run build
npm test
npm run privacy -- --binary
node scripts/dependency-audit.mjs
npm audit --omit=devnode tests/opencode-smoke.mjs uses installed OpenCode with an isolated home and a
local deterministic provider. It verifies integration, not live model judgment
quality. Set ARC_TEST_SCENARIO=stop for native cancellation testing.
The current website is maintained separately in Pumkin4534/arc-mcp-site.
The site/ directory here is the earlier private-preview design. No GitHub Pages
deployment is enabled for this MCP repository.
To remove Arc, delete only mcp.arc from the selected config and the installed
Arc files: arc-mcp/, plugins/arc-opencode.js, agents/arc-reviewer.md,
skills/arc-workflow/, and commands/{arc,arc-status,arc-continue}.md. Then restart
OpenCode. Keep your evidence directory if you want to retain workflow history.
MIT licensed. See LICENSE.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Plan and verify evidence-aware research workflows with specialist execution receipts.
Create, validate, and manage agent-powered workflows with OutcomeCI.
Pre-execution governance for AI agents. Deterministic PASS/FAIL/REVIEW verdicts, replayable proof.
Preflight, approve, and prove consequential agent actions with signed evidence and x402 tools.
Related MCP Servers
FlicenseNot gradedqualityDmaintenanceEnables orchestration of autonomous coding agents (Claude Code, Cursor, etc.) through an objective-native planning board with hash-chained audit trail. Humans define outcomes, agents claim and execute tasks via MCP.7-- AlicenseNot gradedqualityAmaintenanceEnables evidence-gated, multi-session AI coding runs with plan-build-ship state management, coordinating Claude Code and Codex native agents.20 npmMIT
- FlicenseBqualityBmaintenanceEnables AI coding assistants to run a machine-verified DESIGN→PLAN→EXECUTE→VERIFY→COMPLETE workflow with human approval gates, state integrity checks, and DAG task scheduling.7-
- AlicenseNot gradedqualityBmaintenanceEnables bounded, multi-phase coding workflows inside Codex that plan, implement, independently review, repair, and verify repository changes, with an inline dashboard for inspecting runs.MIT