Other MCP clients

Every client below shares the same flow: start the server with npx @agentmemory/agentmemory, wire the client, restart it, then verify with one prompt and the viewer at http://localhost:3113 or curl -X POST http://localhost:3111/agentmemory/smart-search -H 'Content-Type: application/json' -d '{"query": "test"}'. The first six have automated connect adapters that back up the target file, merge the entry, and verify the write. The last three are manual installs; connect prints the exact instructions instead of mutating files.

Automated adapters

npx @agentmemory/agentmemory connect cline

Detects ~/.cline/ and writes the standard block to ~/.cline/mcp.json (the CLI surface):

~/.cline/mcp.json

{
  "mcpServers": {
    "agentmemory": {
      "command": "npx",
      "args": ["-y", "@agentmemory/mcp"],
      "env": {
        "AGENTMEMORY_URL": "${AGENTMEMORY_URL:-http://localhost:3111}",
        "AGENTMEMORY_SECRET": "${AGENTMEMORY_SECRET:-}",
        "AGENTMEMORY_TOOLS": "${AGENTMEMORY_TOOLS:-all}"
      }
    }
  }
}

VS Code extension users paste the same block via Cline Settings, MCP Servers, Edit JSON. The guideline lands as a project rule at .clinerules/agentmemory.md (no frontmatter, always active).

npx @agentmemory/agentmemory connect continue

Continue v1+ prefers ~/.continue/config.yaml, and the adapter branches on what exists:

config.yaml exists: the adapter refuses to rewrite your YAML (preserving comments and anchors safely needs a parser it does not ship) and prints the exact block to merge.
only legacy config.json exists: the entry is merged in place. Continue uses array form, {"name": "agentmemory", "command": "npx", ...} inside a mcpServers list, not an object.
neither exists: a fresh config.yaml is created:

~/.continue/config.yaml

mcpServers:
  - name: agentmemory
    command: npx
    args:
      - "-y"
      - "@agentmemory/mcp"
    env:
      AGENTMEMORY_URL: "${AGENTMEMORY_URL:-http://localhost:3111}"
      AGENTMEMORY_SECRET: "${AGENTMEMORY_SECRET:-}"
      AGENTMEMORY_TOOLS: "${AGENTMEMORY_TOOLS:-all}"

Guideline: project rule at .continue/rules/00-agentmemory.md (Continue does not auto-read AGENTS.md).

npx @agentmemory/agentmemory connect kiro

Detects ~/.kiro/ and writes the standard block to the user-level ~/.kiro/settings/mcp.json. Workspace overrides live in .kiro/settings/mcp.json next to your code. Guideline: a steering file at ~/.kiro/steering/agentmemory.md with inclusion: always frontmatter.

npx @agentmemory/agentmemory connect qwen

Detects ~/.qwen/ and writes the standard block into ~/.qwen/settings.json (Qwen’s mcpServers shape matches Claude Code’s). Qwen’s hook payload is field-compatible with Claude Code, so the bundled 12-hook scripts work unmodified when wired through the hooks section of the same settings.json. Guideline: marked block in ~/.qwen/QWEN.md.

npx @agentmemory/agentmemory connect dsh

Detects the Harness home (DSH_HOME, defaulting to ~/.dsh) and appends one row to the home-level cordis.patch.yml, the machine-local patch layer every Harness profile loads:

$DSH_HOME/cordis.patch.yml

- insert:
    - id: agentmemory
      name: '@deepseek-ai/dsh-mcp-client'
      config:
        transport: stdio
        serverName: agentmemory
        command: npx
        args: ['-y', '@agentmemory/mcp']
        env:
          AGENTMEMORY_URL: http://localhost:3111

The bundled mcp-client plugin registers the tools as mcp__agentmemory__<tool> before the first turn. Harness env blocks take literal strings (no shell interpolation), so edit the URL directly if your engine binds elsewhere. Re-running is idempotent; --force replaces the rows.For auto-capture, add --with-hooks:

npx @agentmemory/agentmemory connect dsh --with-hooks

This writes agentmemory’s Claude Code shaped hook manifest to $DSH_HOME/agentmemory.hooks.json (absolute script paths, no env expansion needed) and adds a second patch row for @deepseek-ai/dsh-hooks-claude-code, Harness’s first-party Claude Code hook bridge. Prompts and tool calls then flow into memory automatically on the bridge’s supported points: SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, and Stop. The manifest’s PreCompact entry is outside the bridge’s supported subset and is skipped with a warning. The bridge runs hooks through Harness’s shell service (subprocess + bash rows), which the standard dsh-base bundle provides; a hand-rolled minimal composition must include them or the bridge stays pending.Both rows are verified against a live Harness run from source: the mcp-client row registers the four memory tools and executes them through the real tool pipeline, and the hook bridge delivers the Claude Code payload to the bundled capture scripts, landing tool observations in the daemon with tool-channel provenance.

npx @agentmemory/agentmemory connect warp

Detects ~/.warp/ and writes the standard block to ~/.warp/.mcp.json. Warp also auto-discovers skills from .claude/skills/, so with the Claude Code plugin installed the 8 agentmemory action skills appear in Warp’s slash-command palette. Guideline: marked block in the project’s AGENTS.md (Warp’s global Rules are UI-managed, so there is no global file to write).

npx @agentmemory/agentmemory connect zed

Zed keys MCP servers under context_servers, not mcpServers, in ~/.config/zed/settings.json (Zed uses the XDG path even on macOS):

~/.config/zed/settings.json

{
  "context_servers": {
    "agentmemory": {
      "command": "npx",
      "args": ["-y", "@agentmemory/mcp"],
      "env": {
        "AGENTMEMORY_URL": "${AGENTMEMORY_URL:-http://localhost:3111}",
        "AGENTMEMORY_SECRET": "${AGENTMEMORY_SECRET:-}",
        "AGENTMEMORY_TOOLS": "${AGENTMEMORY_TOOLS:-all}"
      }
    }
  }
}

Guideline: marked block in ~/.config/zed/AGENTS.md.

Manual installs

Hermes Agent
pi
OpenHuman

connect hermes detects ~/.hermes/ but does not rewrite YAML; it prints this block to add to ~/.hermes/config.yaml:

~/.hermes/config.yaml

mcp_servers:
  agentmemory:
    command: npx
    args: ["-y", "@agentmemory/mcp"]

memory:
  provider: agentmemory

For deeper integration, copy the Python plugin:

cp -r integrations/hermes ~/.hermes/plugins/agentmemory

The plugin hooks six points in the agent loop: prefetch() injects memories before each LLM call, sync_turn() captures every turn, on_session_end() marks sessions for summarization, on_pre_compress() re-injects context before compaction, on_memory_write() mirrors MEMORY.md writes, and system_prompt_block() injects the project profile at session start. It reads ~/.agentmemory/.env at import time to fill missing AGENTMEMORY_URL and AGENTMEMORY_SECRET values; shell exports take precedence. Set AGENTMEMORY_REQUIRE_HTTPS=1 to refuse sending the bearer token over plaintext HTTP to a non-loopback host.Validate the install with Hermes’s own tooling: hermes plugins doctor ~/.hermes/plugins/agentmemory checks the manifest, import, and runtime contracts, and hermes memory status should show the provider as installed and available. Verified against Hermes v0.20.1.

pi uses its extension API, not MCP, so it hooks the agent lifecycle directly against the REST API on 3111. One command installs it:

npx @agentmemory/agentmemory connect pi

The bundled extension lands in ~/.pi/agent/extensions/agentmemory/, which pi auto-discovers — no settings.json edit. A running pi picks it up with /reload. Re-running is a no-op while the copy is current; a stale copy refreshes with a backup.From a repo checkout, integrations/pi is also a pi package, so pi install ./integrations/pi works and pi install manages it like any other package.The extension adds memory_health, memory_search, and memory_save tools, recall on before_agent_start, capture on agent_end, and the /agentmemory-status command. Verify inside pi with /agentmemory-status; it should report agentmemory healthy.

connect openhuman detects ~/.openhuman/ and reports a stub: no integrations/openhuman/ folder exists in the repo yet. OpenHuman is a Memory-trait host; the expected wiring is its Memory trait implemented against the REST proxy at http://localhost:3111. Automation lands once that integration folder does.

Any other MCP client

The standalone stdio server works with any MCP host:

npx @agentmemory/mcp

Point the host’s config at that command with AGENTMEMORY_URL in the env. Sandboxed clients (Flatpak, Snap, restrictive containers) that cannot reach the host’s localhost should also set "AGENTMEMORY_FORCE_PROXY": "1" and point AGENTMEMORY_URL at a route the sandbox can reach. Adapter sources: src/cli/connect/{cline,continue,dsh,kiro,qwen,warp,zed,hermes,pi,openhuman}.ts. Plugins: integrations/hermes/, integrations/pi/.

Get started

Concepts

Connect

Integrations

Workflows

Automated adapters

Manual installs

Any other MCP client

​Automated adapters

​Manual installs

​Any other MCP client

Automated adapters

Manual installs

Any other MCP client