内容
# agentprism
**A universal MCP server that exposes coding agents as background subagents behind a single, unified tool interface.**
agentprism lets one AI agent orchestrate other AI agents. Drop it into your MCP client (Claude Code, Cursor, Continue, …) and gain nine tools — `agent_run`, `agent_spawn`, `agent_resume`, `agent_wait`, `agent_status`, `agent_list`, `agent_kill`, `agent_models`, `agent_providers` — that drive any supported coding agent through its native CLI. Run several in parallel, hand off tasks between them, or use a cheaper/local model as a worker while a more expensive planner directs it.
**8 providers out of the box:** GitHub Copilot, Claude Code, Codex, Gemini CLI, Ollama (local), OpenCode, Aider.
## Installation
**Recommended — no install required ([uvx](https://docs.astral.sh/uv/)):**
```bash
uvx agentprism
```
**Or install permanently:**
```bash
pip install agentprism
# or: uv tool install agentprism
```
**Or from source:**
```bash
git clone https://github.com/StefanMaron/agentprism
cd agentprism
pip install -e .
```
You need at least one supported CLI installed and authenticated:
| Provider | CLI install | Auth command |
|--------------|---------------------------------------------------------------|-------------------------------|
| Copilot | [GitHub Copilot CLI][copilot-cli] | `copilot login` |
| Claude Code | [claude.ai/code][claude-cli] | `claude` → `/login` |
| Codex | [github.com/openai/codex][codex-cli] | `codex login` |
| Gemini | `npm install -g @google/gemini-cli` | `gemini auth login` |
| Ollama | [ollama.com][ollama] | pull a model: `ollama pull qwen2.5-coder:14b` |
| OpenCode | `npm install -g opencode-ai` | `opencode providers login` |
| Aider | `uv tool install --python 3.12 aider-chat` | set `ANTHROPIC_API_KEY` or use Ollama |
[copilot-cli]: https://docs.github.com/en/copilot/github-copilot-in-the-cli
[claude-cli]: https://docs.anthropic.com/en/docs/claude-code
[codex-cli]: https://github.com/openai/codex
[ollama]: https://ollama.com
## Usage with Claude Code
Register globally (recommended):
```bash
claude mcp add agentprism --scope user -- agentprism
# or via uvx (no install needed):
claude mcp add agentprism --scope user -- uvx agentprism
```
Or add manually to `~/.claude/mcp.json`:
```json
{
"mcpServers": {
"agentprism": {
"command": "agentprism",
"type": "stdio"
}
}
}
```
Restart Claude Code. The nine `agent_*` tools appear. Example prompt:
> Call `agent_providers` to see what's available, then use `agent_run` to let Copilot write a Python script that prints prime numbers up to 100 in `/tmp/playground`.
## Dashboard
Every running agentprism instance auto-starts an HTTP API on a random free port and registers itself in `~/.agentprism/{pid}.json`. A standalone dashboard discovers and aggregates all running instances:
```bash
agentprism dashboard # default port 7070
agentprism dashboard --port 8080
```
Open `http://localhost:7070` to see every active session across every project, grouped by working directory, with live streaming output per session.
## Tool reference
| Tool | Args | Returns |
|--------------------|---------------------------------------------------|--------------------------------------------------|
| `agent_providers` | — | which providers are installed + authenticated |
| `agent_models` | `provider?` | model ids + cost multipliers |
| `agent_run` | `task`, `cwd`, `provider?`, `model?`, `timeout?` | output — one-shot, blocks until done, auto-cleans |
| `agent_spawn` | `task`, `cwd`, `provider?`, `model?`, `mode?` | `session_id` — non-blocking, persistent |
| `agent_resume` | `session_id`, `message` | start a new turn on an existing session — must not be `working` |
| `agent_status` | `session_id` | status + git delta + activity info |
| `agent_wait` | `session_id`, `timeout_seconds?` | accumulated output + git delta (blocks) |
| `agent_list` | — | all active sessions |
| `agent_kill` | `session_id` | terminates the subprocess |
**`agent_status`** returns `process_alive`, `last_activity_seconds_ago`, `uptime_seconds`, `new_commits`, and `working_tree_changes` — enough to decide whether a session is stuck without calling `agent_wait`.
**`agent_wait` / `agent_run`** include `new_commits` and `working_tree_changes` in the result — no need to run `git log` or `git status` separately. Each call is capped server-side at `AGENTPRISM_WAIT_CAP_SECONDS` (default 60); if the worker is still running when the cap fires, the response is `{status: "still_running", ...}` and the caller should call `agent_wait` again. The worker is left alive across the gap. This prevents long jobs from outliving the MCP host's per-tool-call timeout.
**Session persistence & recovery.** Every spawn writes `~/.agentprism/sessions/{session_id}.json` capturing the worker PID and PGID. If agentprism dies (host crash, restart, MCP disconnect), workers keep running — they're in their own process group, not children of the host. The next agentprism instance scans those files on startup and rehydrates any session whose worker is still alive, so `agent_list` / `agent_status` / `agent_kill` keep working across restarts. Recovered sessions show `recovered: true` in `agent_list`. `agent_resume` is unavailable on recovered sessions (the original stdin pipe is gone) — kill and respawn if you need to redirect.
**`mode`** values (Copilot / Claude Code): `agent` (default), `plan`, `autopilot`
> **Note:** Background agents always run in **yolo mode** — they auto-approve all file edits and shell commands without prompting. This is intentional: subagents are unattended by design. Review their output (via `agent_wait` or the dashboard) before merging or running produced code.
## Provider guide
### Free — no quota
| Provider | `provider=` | Model example | Notes |
|-----------|-------------|----------------------------------|-----------------------------------------|
| Copilot | `copilot` | `gpt-4.1`, `gpt-5-mini` | 0x quota, requires Copilot subscription |
| Ollama | `ollama` | `qwen2.5-coder:14b-8k` | Local GPU — best for Q&A, single-file |
| OpenCode | `opencode` | `opencode/big-pickle` | Free bundled models, no API key |
### Free — quota-limited
| Provider | `provider=` | Model | Quota |
|----------|-------------|-------------------------|-----------|
| Gemini | `gemini` | `gemini-2.5-flash` | ~1000/day |
| Gemini | `gemini` | `gemini-2.0-flash` | Higher |
### Paid
| Provider | `provider=` | Model | Cost | Notes |
|-------------|-------------|-------------------------|--------|-------------------------------|
| Copilot | `copilot` | `claude-haiku-4.5` | 0.33x | Cheapest Claude |
| Copilot | `copilot` | `gpt-5.4-mini` | 0.33x | Cheapest GPT |
| Copilot | `copilot` | `claude-sonnet-4.6` | 1x | **Default** — most tasks |
| Copilot | `copilot` | `claude-opus-4.7` | 7.5x | Deep reasoning |
| Claude Code | `claude` | any Claude model | varies | Direct Anthropic billing |
| Codex | `codex` | any OpenAI model | varies | Direct OpenAI billing |
| Aider | `aider` | `qwen2.5-coder:14b-8k` | 0x | Local Ollama — git repo req'd |
### Local LLM notes
Ollama and Aider use your local GPU. A 16 GB GPU (e.g. RTX 4080) runs `qwen2.5-coder:14b` well for Q&A and single-file edits (~84 tok/s at 8K context). For reliable multi-file editing, 32 GB+ VRAM (RTX 5090 or workstation GPU) is recommended so 70B quantized models fit.
Create a custom 8K-context modelfile if needed:
```
FROM qwen2.5-coder:14b
PARAMETER num_ctx 8192
```
```bash
ollama create qwen2.5-coder:14b-8k -f Modelfile
```
For OpenCode + Ollama, add to `~/.config/opencode/config.json`:
```json
{
"provider": {
"ollama": {
"npm": "@ai-sdk/openai-compatible",
"options": { "baseURL": "http://localhost:11434/v1" },
"models": {
"qwen2.5-coder:14b-8k": { "name": "Qwen2.5 Coder 14B (8K)" }
}
}
}
}
```
## Quota handling
When a provider hits its quota limit, `agent_run` / `agent_wait` return a structured error instead of a crash:
```json
{
"error": "quota_exceeded",
"provider": "gemini",
"model": "gemini-2.5-flash",
"suggestion": "Try provider='ollama', model='qwen2.5-coder:14b-8k' for free local inference..."
}
```
## Helping your agent know when to delegate
Add a snippet to your project's `CLAUDE.md` or `AGENTS.md`:
```markdown
## Delegation with agentprism
You have access to the agentprism MCP server. Delegate coding tasks to external
agents instead of doing the work yourself to preserve context window.
Trigger conditions:
- User says "let Copilot handle", "delegate", "offload to an agent"
- Task is large/mechanical and would burn significant context
- Multiple tasks can run in parallel
Quick patterns:
- One-shot: agent_run(task, cwd)
- Parallel workers: multiple agent_spawn calls, then agent_wait each
- Follow-up turns: agent_spawn → agent_wait → agent_resume → agent_wait → agent_kill
- Free brainstorm: agent_run(task, cwd, provider="ollama", model="qwen2.5:14b")
- Free Gemini: agent_run(task, cwd, provider="gemini", model="gemini-2.5-flash")
Default provider is copilot (claude-sonnet-4.6, 1x cost). Call agent_providers
to check what's available on this machine.
```
## Push notifications
When a worker finishes, agentprism proactively notifies the orchestrating MCP client.
If the client advertised the `sampling` capability (Claude Code does), agentprism sends a `sampling/createMessage` wake-up with the session summary so the orchestrator can immediately act. Falls back to `notifications/message` for clients that don't support sampling.
## Architecture
```
Claude session A Claude session B agentprism dashboard
(project X) (project Y) (standalone, any terminal)
│ │ │
▼ ▼ │
agentprism (stdio) agentprism (stdio) │
SessionRegistry SessionRegistry │
HTTP API :auto ──────────────────────────────────────────────►│ fan-out
│ writes │ writes reads │
▼ ▼ all ▼
~/.agentprism/{pidA}.json ~/.agentprism/{pidB}.json ──► grouped by project
│ │ http://localhost:7070
▼ ▼
CopilotAdapter ClaudeCodeAdapter
GeminiAdapter OllamaAdapter
AiderAdapter OpenCodeAdapter
│ │
▼ ▼
copilot -p --yolo claude --output-format stream-json
(subprocess) (subprocess)
```
Sessions are fully isolated per Claude session — no cross-session interference. The standalone dashboard is read-only and discovers instances via `~/.agentprism/{pid}.json` lockfiles.
## Configuration
Environment variables:
| Variable | Default | Purpose |
|----------------------------|--------------|------------------------------------------------|
| `AGENTPRISM_LOG_LEVEL` | `INFO` | Python logging level (logs go to stderr) |
| `AGENTPRISM_DEFAULT_PROVIDER` | `copilot` | Provider used when `provider` arg is omitted |
| `AGENTPRISM_WAIT_CAP_SECONDS` | `60` | Server-side cap on each `agent_wait`/`agent_run` call. When the cap fires before the worker finishes, the response has `status: "still_running"` and the caller should call `agent_wait` again. Avoids tearing down the MCP stdio channel on long jobs. |
| `AGENTPRISM_COPILOT_BIN` | `copilot` | Path to the `copilot` binary |
| `AGENTPRISM_CLAUDE_BIN` | `claude` | Path to the `claude` binary |
| `AGENTPRISM_CODEX_BIN` | `codex` | Path to the `codex` binary |
| `AGENTPRISM_GEMINI_BIN` | `gemini` | Path to the `gemini` binary |
| `AGENTPRISM_OPENCODE_BIN` | `opencode` | Path to the `opencode` binary |
| `AGENTPRISM_AIDER_BIN` | `aider` | Path to the `aider` binary |
| `OLLAMA_API_BASE` | `http://localhost:11434` | Ollama endpoint |
## Development
```bash
git clone https://github.com/StefanMaron/agentprism
cd agentprism
pip install -e ".[dev]"
ruff check .
pytest
```
See [CLAUDE.md](CLAUDE.md) for architecture notes and contributor guidance.
## License
MIT
MCP 配置
以下是此 MCP Server 的配置信息，可直接复制到 Cursor 或其他 MCP 客户端中使用。
mcp.json
连接信息
猜你喜欢
暂无推荐服务器