# Buckley

> A tool-first AI agent harness for durable repository work across terminal, browser, ACP, LSP, and CI.

- **Type:** MCP server
- **Install:** `agentstack add mcp-odvcencio-buckley`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [odvcencio](https://agentstack.voostack.com/s/odvcencio)
- **Installs:** 0
- **Category:** [Developer Tools](https://agentstack.voostack.com/c/developer-tools)
- **Latest version:** 0.1.0
- **License:** MIT
- **Upstream author:** [odvcencio](https://github.com/odvcencio)
- **Source:** https://github.com/odvcencio/buckley
- **Website:** https://m31labs.dev/buckley

## Install

```sh
agentstack add mcp-odvcencio-buckley
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# Buckley

[](https://pkg.go.dev/m31labs.dev/buckley)
[](https://github.com/odvcencio/buckley/actions/workflows/ci.yml)
[](https://github.com/odvcencio/buckley/actions/workflows/release.yml)

Buckley is a tool-first AI agent harness for serious repository work.

It combines resumable sessions, Arbiter-governed model and tool selection, Claude-style repository instructions, and multiple operator surfaces: terminal, browser, one-shot, ACP, and LSP.

## Agent Skill

Agents working with Buckley should use the [using-buckley](https://github.com/odvcencio/m31labs-skills/blob/main/skills/using-buckley/SKILL.md) skill.

## Why Buckley

- One shared runtime powers TUI chat, `buckley -p`, headless sessions, ACP editor flows, and browser control.
- Arbiter governs planning/execution/review model routing, reasoning effort, tool pools, timeouts, approvals, and escalation.
- Runtime prompt assembly automatically pulls in `AGENTS.md`, `CLAUDE.md`, `.claude/instructions.md`, project context, and active skills.
- Tool use is first-class. Buckley ships with 38 built-in tool entry points plus skills, plugins, telemetry, approvals, and SQLite-backed persistence.
- Sessions survive crashes. Plans, approvals, telemetry, and artifacts stay resumable.

## Quick Start

```bash
go install m31labs.dev/buckley/cmd/buckley@v0.4.0
export OPENROUTER_API_KEY="your-key"
buckley
```

OpenAI, Anthropic, Google, and Ollama are also supported. OpenRouter is the default path when available.

## Interfaces

| Surface | Command | Use case |
| --- | --- | --- |
| TUI | `buckley` | Interactive coding with approvals, streaming, history, and highlighted code |
| One-shot | `buckley -p "inspect this repo and fix the failing tests"` | Fast task execution from the terminal |
| Browser UI | `buckley serve --browser` | Mission Control, approvals, and remote session control |
| ACP agent | `buckley acp` | Editor agent for ACP-compatible clients |
| LSP bridge | `buckley lsp` | LSP editor integration on stdio |

## Core Workflow

```bash
buckley plan "add auth" "support email/password login"
buckley execute 
buckley review
buckley review-pr 123
buckley review-pr 123 --model codex/auto
```

The planner, builder, review, and runtime layers share the same governance stack, persistence, and tool registry.

## Notable Capabilities

- Fully integrated Arbiter runtime across one-shot, TUI, ACP, and headless execution paths.
- Shared runtime prompt assembly for repo instructions, project context, working directory, and skills.
- Governed tool exposure with role/task-aware pool filtering before tool calls are shown to the model.
- Anthropic tool calling and tool-result round-tripping, so Claude-class models can participate in the same tool loop.
- Better model resolution for routed raw IDs such as unqualified Claude model names.

## Commands People Actually Use

```bash
buckley commit
buckley pr
buckley review
buckley review-pr 123
buckley hunt
buckley dream
buckley experiment run "compare-routing" -m z-ai/glm-5.2 -m moonshotai/kimi-k2.7-code -p "Implement feature X"
```

`buckley review` examines one immutable snapshot of the selected local scope. Untracked files are excluded by default; repeat `--include-untracked path/to/new.go` to explicitly allowlist worktree files that may be sent to the model, after inspecting them for secrets. Native Codex verification receives a self-contained copy of only the captured commit plus patch; its JSONL command events must prove successful, classifiable build and test runs that cover the changed source paths. Use `--model codex/auto` to select Luna with xhigh reasoning for focused changes, Terra with medium reasoning for standard changes, and Sol with medium reasoning for broad or project reviews. An exact model override remains fixed. API-model inspection tools are rooted to an independently materialized copy of that snapshot and can run only deterministic build/test/check plans in an OS-enforced, read-only-source sandbox. `buckley review-pr` adds the remote PR diff, CI state, submitted reviews, and unresolved inline feedback; run it from a checkout whose `HEAD` is the PR head so verification is pinned to the same revision. API approvals require actual successful verification tool calls—not prose claiming PASS—and build/test evidence must use one applicable toolchain and cover every changed source package. Branch and PR approvals require passing verification, complete evidence, an explicit disposition for every supplied feedback ID, and an independent approval-critic pass. `buckley review -project` is an advisory architecture assessment and cannot issue an approval verdict. `buckley commit` and `buckley pr` use transparent tool-first workflows rather than opaque text-only prompting. In the TUI, fenced code blocks use Tree-sitter highlighting when the language grammar is available. Use `Shift+Enter` to add a line without sending and `Alt+C` to copy the latest code block. Conversation projection searches for the largest compacted history that fits the provider request budget.

## Configuration

Configuration is layered:

```text
~/.buckley/config.yaml
./.buckley/config.yaml
environment variables
```

Minimal setup:

```yaml
providers:
  openrouter:
    api_key: ${OPENROUTER_API_KEY}
```

Buckley supports separate planning, execution, and review models:

```yaml
models:
  planning: z-ai/glm-5.2
  execution: z-ai/glm-5.2
  review: moonshotai/kimi-k2.7-code
  reasoning: xhigh
```

## Skills And Instructions

Buckley ships with 8 bundled skills, including planning, code review, systematic debugging, refactoring, test-driven development, API design, git workflow, and creative writing.

Skills layer on top of repository instructions. If a repo already uses `AGENTS.md` or `CLAUDE.md`, Buckley consumes them automatically and applies the same guidance across terminal, headless, and editor sessions.

## Development

```bash
./scripts/test.sh
go build ./cmd/buckley
```

Primary repo docs live under [`docs/`](docs/).

## License

MIT. See [LICENSE](LICENSE).

## Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [odvcencio](https://github.com/odvcencio)
- **Source:** [odvcencio/buckley](https://github.com/odvcencio/buckley)
- **License:** MIT
- **Homepage:** https://m31labs.dev/buckley

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/mcp-odvcencio-buckley
- Seller: https://agentstack.voostack.com/s/odvcencio
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.