Skip to main content
Glama

AgentBridge MCP

A stdio TypeScript MCP server that delegates bounded coding tasks to a server-side agent loop. It persists tasks and conversations in SQLite, exposes compact task receipts, and keeps file access and command execution inside a guarded workspace.

Current version: 0.1.1. See CHANGELOG.md for release history; every change updates both the version and changelog.

Install and run

npm.cmd install
npm.cmd run build
$env:AGENTBRIDGE_WORKSPACE = "D:\workplace\my-project"
$env:AGENTBRIDGE_API_KEY = "<server-side key>"
$env:AGENTBRIDGE_API_BASE = "https://api.openai.com/v1" # OpenAI-compatible endpoint
$env:AGENTBRIDGE_MODEL = "gpt-4.1-mini"
npm.cmd start

MCP client configuration:

{
  "mcpServers": {
    "agentbridge": {
      "command": "node",
      "args": ["D:/workplace/2026/AgentBridge MCP/dist/index.js"],
      "env": {
        "AGENTBRIDGE_WORKSPACE": "D:/workplace/my-project",
        "AGENTBRIDGE_API_KEY": "${AGENTBRIDGE_API_KEY}",
        "AGENTBRIDGE_MODEL": "gpt-4.1-mini"
      }
    }
  }
}

Related MCP server: PatchWarden

MCP tools

delegate_task, get_task, continue_task, get_diff, and cancel_task are public. The agent alone can call workspace_list, workspace_search, workspace_read, workspace_patch, and workspace_command.

Security model

  • Files must resolve through configured workspace roots; .git, symlinks, and common secret files are blocked.

  • Patches are a single-file unified diff with a required SHA-256 precondition. File creation, deletion, and renames are disabled. Each write has before/after audit hashes.

  • Commands are spawned without a shell, allow-listed (npm, npx, node, git), time-limited, and output-capped.

  • API keys are read only from server environment variables and redacted from task errors. MCP receipts omit full source, prompt history, and raw tool logs.

Limits

The OpenAI-compatible adapter supports chat-completions tool calls. The Codex CLI adapter is reserved but deliberately not implemented; no arbitrary shell execution is provided. Set AGENTBRIDGE_ALLOWED_ROOTS (semicolon-separated) to constrain the server further. State is written under .agentbridge/ unless AGENTBRIDGE_STATE_DIR is provided.

Verification

npm.cmd run check
npm.cmd test

Available Tools

5 tools
cancel_taskA

Cancel a task. It can no longer be resumed.

Inputs: task_id (required).

continue_taskA

Resume a paused or failed task using its persisted conversation.

Inputs: task_id (required), prompt.

delegate_taskB

Create an asynchronous secure coding task. Returns only its task and conversation handles.

Inputs: prompt (required), workspace (required), allowed_paths.

get_diffB

Get a small audited change record for a task or file. Full source and tool logs are excluded.

Inputs: task_id (required), path.

get_taskA

Get a compact task receipt; does not expose model transcripts or tool logs.

Inputs: task_id (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.9/5.0

    Scored across 5 tools

    Disambiguation5/5

    Each tool targets a distinct operation: retrieving a task receipt, resuming a task, fetching a diff, canceling, and delegating. There is no overlap or ambiguity between these actions.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun pattern (get_task, continue_task, get_diff, cancel_task, delegate_task) using lowercase snake_case throughout.

    Tool Count5/5

    Five tools is a well-scoped size for a task orchestration server, covering the essential lifecycle (create, read, resume, cancel, diff) without bloat or redundancy.

    Completeness4/5

    The tool set covers the full task lifecycle including creation, retrieval, resumption, cancellation, and change auditing. A minor gap is the lack of a list/query tool, but the design using handles makes it acceptable.

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      MCP server that lets multiple coding-agent sessions on the same machine discover each other and collaborate through a shared SQLite database.
      1
      MIT
    • F
      license
      Not graded
      quality
      D
      maintenance
      SQLite-backed MCP server for Claude Code session persistence and multi-agent coordination. Provides tools for session management, event logging, decision tracking, file locking, agent registry, and plan tracking.
      -
    • A
      license
      A
      quality
      C
      maintenance
      MCP server providing structured continuity memory for AI coding agents, tracking decisions, open loops, and session state in local SQLite. Enables agents to resume work from verified state across sessions without replaying transcripts.
      17
      Apache 2.0