MCP shim: --tools all returns only 7 tools due to IMPLEMENTED_TOOLS hardcode

[cl0ckt0wer]

Summary

agentmemory mcp --tools all (and AGENTMEMORY_TOOLS=all env var) silently returns only 7 tools instead of 51. The issue is in src/mcp/standalone.ts — even though getVisibleTools() correctly returns all 51 when mode is "all", the result is filtered through a hardcoded IMPLEMENTED_TOOLS set of 7 tool names.

Root cause

src/mcp/standalone.ts:

const IMPLEMENTED_TOOLS = new Set([
  "memory_save", "memory_recall", "memory_smart_search",
  "memory_sessions", "memory_export", "memory_audit",
  "memory_governance_delete",
]);

tools/list returns: getVisibleTools().filter((t) => IMPLEMENTED_TOOLS.has(t.name))

Even with --tools all, getVisibleTools() returns all 51, then the filter reduces it to 7.

Impact

Every non-Claude-Code user (OpenCode, Cursor, Gemini CLI, Cline, etc.) is affected. Users get 7 tools and have no path to the remaining 44 (actions, graphs, signals, sentinels, slots, lessons, etc.) without writing a custom HTTP→stdio bridge.

Proposed fix

When in proxy mode (detected server via AGENTMEMORY_URL), delegate tools/list to the server: GET <base>/agentmemory/mcp/tools and return its response directly, bypassing the local filter. The server already has all 51 tools registered.

Environment

• agentmemory: v0.9.4 (Docker, iii-engine 0.11.6)
• Client: OpenCode (stdio MCP)
• Workaround: custom 22-line Node.js bridge proxying stdio → HTTP on Docker worker

[rohitg00]

Thanks @cl0ckt0wer — root cause is exactly right. Confirmed:

• src/mcp/standalone.ts hardcodes a 7-name IMPLEMENTED_TOOLS set
• tools/list filters getVisibleTools() through that set even when AGENTMEMORY_TOOLS=all is configured
• Result: every non-Claude-Code MCP client gets 7 tools regardless of the config flag

Fix is in #238 — switches tools/list and validate() to proxy-aware behavior: when running against a reachable server (AGENTMEMORY_URL set + livez ok), delegate to the server's tool list (gets all 51), with a fallback to the local 7-tool list when the server is unreachable. Has tests covering the proxy path including auth-header passthrough.

I've already approved #238 — holding for @rohitg00 to land. Once it does, will close this with the merge.

Good catch on the impact framing too — every non-Claude-Code user (OpenCode, Cursor, Gemini CLI, Cline) was hitting this silently.

[jcalfee]

◇ iii-engine is ready
[agentmemory] Starting worker v0.9.6...

[rohitg00]

@jcalfee what should I see here?

[jcalfee]

It appear that version 0.9.6 didn't work because I'm seeing only 4 tools. If I'm reading the github comments right then it would have been fixed in version 0.9.6.

I'm running VS Codeium in Flatpak with the MCP server in RooCode.

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

It seems that the RooCode MCP startup is able to escape the flat pack though because the place where I find it running the AgentMemory MCP server is on the host. So that seems to go well.

So I check on the host and see this version:

$ npm exec @agentmemory/mcp --version
10.9.7

And I get these four tools whether or not I have the regular process running on the host, which was very confusing until I figured out or remembered that you've got a Docker container. So that makes sense that it's coming from the Docker container:

$ npx @agentmemory/agentmemory
┌  agentmemory
│
◆  iii-engine is running
[agentmemory] Starting worker v0.9.6...
...

I'm not sure how to check the version inside of the Docker container which is actually pod man on my machine because I can't find bash or sh inside but I assumed that this should be up to date and the bug would be fixed right?

agentmemory-iii-engine-1

[rohitg00]

@jcalfee thanks for the detail — the 4-tool count is the smoking gun. Two things going on:

1. Version check is misleading
npm exec @agentmemory/mcp --version reports 10.9.7 because --version is consumed by npm itself (your npm CLI version), not the package. To check the package version:

npx -y @agentmemory/mcp@latest 2>&1 | head -3
# or
npm view @agentmemory/mcp version

Both should show 0.9.6.

2. The 4 tools come from the container, not the shim
The shim flow is:

• probe http://localhost:3111/agentmemory/livez (500ms)
• if up → GET /agentmemory/mcp/tools from the server → return whatever the server lists (8 essentials by default, 51 when AGENTMEMORY_TOOLS=all)
• if down → fall back to local 7-tool IMPLEMENTED_TOOLS

4 ≠ 8 and ≠ 7, so your container is older than v0.9.6 and its /agentmemory/mcp/tools predates the current ESSENTIAL_TOOLS set.

Fix — rebuild the container (swap docker → podman if that's your runtime)

docker compose down
docker rmi agentmemory-iii-engine ghcr.io/rohitg00/agentmemory:latest 2>/dev/null
docker compose pull
docker compose up -d

# verify the server tool count
curl -s http://localhost:3111/agentmemory/mcp/tools | jq '.tools | length'
# expect 8 (default "core") or 51 (with AGENTMEMORY_TOOLS=all)

To get all 51 tools add to your compose env:

environment:
  AGENTMEMORY_TOOLS: all

Then curl ... | jq '.tools | length' should return 51, and RooCode will see the same set via the shim.

If after rebuild you still see 4 tools from the curl, post the curl output and docker compose ps (or podman compose ps) and I'll dig further.

[rohitg00]

Reopening pending @jcalfee confirmation on Docker rebuild — original fix landed in v0.9.6 (#270) but his container appears to predate the change.

[jcalfee]

I don't think it was the rebuild that was required. It was that I had to put all in the tool names here on the host:

AGENTMEMORY_TOOLS=all npx @agentmemory/agentmemory

I did the rebuild anyways because I was still having problems getting that list of tools into the Flatpak and Roo code in VS Codeium:

ADDED: user@s1max:~/.npm/_npx/076f5982d055bff5/node_modules/@agentmemory/agentmemory

$ egrep AGENTMEMORY_TOOLS docker-compose.yml 
      AGENTMEMORY_TOOLS: all

 user@s1max:~/.npm/_npx/076f5982d055bff5/node_modules/@agentmemory/agentmemory
$ docker compose down
[+] down 2/2
 ✔ Container agentmemory-iii-engine-1 Removed                                                                                                                                       0.7s
 ✔ Network agentmemory_default        Removed                                                                                                                                       0.0s
 user@s1max:~/.npm/_npx/076f5982d055bff5/node_modules/@agentmemory/agentmemory
$ docker compose pull
[+] pull 21/21
 ✔ Image iiidev/iii:0.11.2 Pulled                                                                                                                                                   0.9s
 user@s1max:~/.npm/_npx/076f5982d055bff5/node_modules/@agentmemory/agentmemory
$ docker compose up -d
[+] up 2/2
 ✔ Network agentmemory_default        Created                                                                                                                                       0.0s
 ✔ Container agentmemory-iii-engine-1 Started                                                                                                                                       0.1s
 user@s1max:~/.npm/_npx/076f5982d055bff5/node_modules/@agentmemory/agentmemory
$ curl -s http://localhost:3111/agentmemory/mcp/tools | jq '.tools | length'
 user@s1max:~/.npm/_npx/076f5982d055bff5/node_modules/@agentmemory/agentmemory
$ curl -s http://localhost:3111/agentmemory/mcp/tools | jq '.tools | length'
51

Now the issue persists:

 user@s1max:~/.npm/_npx/076f5982d055bff5/node_modules/@agentmemory/agentmemory
$ pgrep -fal agentmemory
466782 npm exec @agentmemory/agentmemory
466797 node /home/user/.npm/_npx/076f5982d055bff5/node_modules/.bin/agentmemory
471625 /usr/bin/conmon --api-version 1 -c 0157bb72948d4a83c4ef28334eb88a89c576a14c4476740c85737d605d59b01d -u 0157bb72948d4a83c4ef28334eb88a89c576a14c4476740c85737d605d59b01d -r /usr/bin/crun -b /home/user/.local/share/containers/storage/overlay-containers/0157bb72948d4a83c4ef28334eb88a89c576a14c4476740c85737d605d59b01d/userdata -p /run/user/1000/containers/overlay-containers/0157bb72948d4a83c4ef28334eb88a89c576a14c4476740c85737d605d59b01d/userdata/pidfile -n agentmemory-iii-engine-1 --exit-dir /run/user/1000/libpod/tmp/exits --persist-dir /run/user/1000/libpod/tmp/persist/0157bb72948d4a83c4ef28334eb88a89c576a14c4476740c85737d605d59b01d --full-attach -s -l journald --log-level info --syslog --runtime-arg --log-format=json --runtime-arg --log --runtime-arg=/run/user/1000/containers/overlay-containers/0157bb72948d4a83c4ef28334eb88a89c576a14c4476740c85737d605d59b01d/userdata/oci-log --conmon-pidfile /run/user/1000/containers/overlay-containers/0157bb72948d4a83c4ef28334eb88a89c576a14c4476740c85737d605d59b01d/userdata/conmon.pid --exit-command /usr/bin/podman --exit-command-arg --root --exit-command-arg /home/user/.local/share/containers/storage --exit-command-arg --runroot --exit-command-arg /run/user/1000/containers --exit-command-arg --log-level --exit-command-arg info --exit-command-arg --cgroup-manager --exit-command-arg systemd --exit-command-arg --tmpdir --exit-command-arg /run/user/1000/libpod/tmp --exit-command-arg --network-config-dir --exit-command-arg  --exit-command-arg --network-backend --exit-command-arg netavark --exit-command-arg --volumepath --exit-command-arg /home/user/.local/share/containers/storage/volumes --exit-command-arg --db-backend --exit-command-arg sqlite --exit-command-arg --transient-store=false --exit-command-arg --hooks-dir --exit-command-arg /usr/share/containers/oci/hooks.d --exit-command-arg --runtime --exit-command-arg crun --exit-command-arg --storage-driver --exit-command-arg overlay --exit-command-arg --events-backend --exit-command-arg journald --exit-command-arg container --exit-command-arg cleanup --exit-command-arg --stopped-only --exit-command-arg 0157bb72948d4a83c4ef28334eb88a89c576a14c4476740c85737d605d59b01d
473013 npm exec @agentmemory/mcp
473014 npm exec @agentmemory/mcp
473072 node /home/user/.npm/_npx/ba4b5775a0ab44e2/node_modules/.bin/agentmemory-mcp
473079 node /home/user/.npm/_npx/ba4b5775a0ab44e2/node_modules/.bin/agentmemory-mcp
473138 npm exec @agentmemory/mcp
473173 node /home/user/.npm/_npx/ba4b5775a0ab44e2/node_modules/.bin/agentmemory-mcp

I did a pkill -9 and docker compose down then restarted the npx host and the MCP server:

 user@s1max:~/.npm/_npx/076f5982d055bff5/node_modules/@agentmemory/agentmemory
$ pgrep -fal agentmemory
476643 npm exec @agentmemory/agentmemory
476815 node /home/user/.npm/_npx/076f5982d055bff5/node_modules/.bin/agentmemory
476889 /usr/bin/conmon --api-version 1 -c a1f11316a57cf95fb49cd5621c8e9ac29d382db6475c3115c042f79f2d1b43a6 -u a1f11316a57cf95fb49cd5621c8e9ac29d382db6475c3115c042f79f2d1b43a6 -r /usr/bin/crun -b /home/user/.local/share/containers/storage/overlay-containers/a1f11316a57cf95fb49cd5621c8e9ac29d382db6475c3115c042f79f2d1b43a6/userdata -p /run/user/1000/containers/overlay-containers/a1f11316a57cf95fb49cd5621c8e9ac29d382db6475c3115c042f79f2d1b43a6/userdata/pidfile -n agentmemory-iii-engine-1 --exit-dir /run/user/1000/libpod/tmp/exits --persist-dir /run/user/1000/libpod/tmp/persist/a1f11316a57cf95fb49cd5621c8e9ac29d382db6475c3115c042f79f2d1b43a6 --full-attach -s -l journald --log-level info --syslog --runtime-arg --log-format=json --runtime-arg --log --runtime-arg=/run/user/1000/containers/overlay-containers/a1f11316a57cf95fb49cd5621c8e9ac29d382db6475c3115c042f79f2d1b43a6/userdata/oci-log --conmon-pidfile /run/user/1000/containers/overlay-containers/a1f11316a57cf95fb49cd5621c8e9ac29d382db6475c3115c042f79f2d1b43a6/userdata/conmon.pid --exit-command /usr/bin/podman --exit-command-arg --root --exit-command-arg /home/user/.local/share/containers/storage --exit-command-arg --runroot --exit-command-arg /run/user/1000/containers --exit-command-arg --log-level --exit-command-arg info --exit-command-arg --cgroup-manager --exit-command-arg systemd --exit-command-arg --tmpdir --exit-command-arg /run/user/1000/libpod/tmp --exit-command-arg --network-config-dir --exit-command-arg  --exit-command-arg --network-backend --exit-command-arg netavark --exit-command-arg --volumepath --exit-command-arg /home/user/.local/share/containers/storage/volumes --exit-command-arg --db-backend --exit-command-arg sqlite --exit-command-arg --transient-store=false --exit-command-arg --hooks-dir --exit-command-arg /usr/share/containers/oci/hooks.d --exit-command-arg --runtime --exit-command-arg crun --exit-command-arg --storage-driver --exit-command-arg overlay --exit-command-arg --events-backend --exit-command-arg journald --exit-command-arg container --exit-command-arg cleanup --exit-command-arg --stopped-only --exit-command-arg a1f11316a57cf95fb49cd5621c8e9ac29d382db6475c3115c042f79f2d1b43a6
477288 npm exec @agentmemory/mcp
477331 npm exec @agentmemory/mcp
477332 npm exec @agentmemory/mcp
477363 node /home/user/.npm/_npx/ba4b5775a0ab44e2/node_modules/.bin/agentmemory-mcp
477384 node /home/user/.npm/_npx/ba4b5775a0ab44e2/node_modules/.bin/agentmemory-mcp
477391 node /home/user/.npm/_npx/ba4b5775a0ab44e2/node_modules/.bin/agentmemory-mcp

Still 51 tools on the host but 7 in Roo Code.

Also, I loose everything when I restart docker and the data dir is empty:

$ find /home/user/.local/share/containers/storage/volumes/agentmemory_iii-data/_data
/home/user/.local/share/containers/storage/volumes/agentmemory_iii-data/_data

The last startup was a bit strange that it had to re-install:

 user@s1max:~/dev/vendor/stt (master)
$ AGENTMEMORY_TOOLS=all npx @agentmemory/agentmemory
Need to install the following packages:
@agentmemory/agentmemory@0.9.6
Ok to proceed? (y) y

┌  agentmemory
which: no iii in (/home/user/.npm/_npx/076f5982d055bff5/node_modules/.bin:/home/user/dev/vendor/stt/node_modules/.bin:/home/user/dev/vendor/node_modules/.bin:/home/user/dev/node_modules/.bin:/home/user/node_modules/.bin:/home/node_modules/.bin:/node_modules/.bin:/usr/lib/node_modules_22/npm/node_modules/@npmcli/run-script/lib/node-gyp-bin:/home/user/.cargo/bin:/home/user/.local/share/mise/installs/http-mise-inpower/1.0.0:/home/user/.local/bin:/home/user/bin:/usr/local/bin:/usr/bin:/run/host/usr/bin:/run/host/sbin:~/dev/ipall/bin:node_modules/.bin:../node_modules/.bin:../../node_modules/.bin:../../../node_modules/.bin:../../../../node_modules/.bin:~/.cargo/bin:~/dev/ipall/bin:node_modules/.bin:../node_modules/.bin:../../node_modules/.bin:../../../node_modules/.bin:../../../../node_modules/.bin:~/.cargo/bin)
│
◇  Docker compose started
│
◇  iii-engine is ready
[agentmemory] Starting worker v0.9.6...
[agentmemory] Engine: ws://localhost:49134
[agentmemory] Provider: openrouter (anthropic/claude-sonnet-4-20250514)
[agentmemory] Embedding provider: openrouter (1536 dims)
[agentmemory] REST API: http://localhost:3111/agentmemory/*
[agentmemory] Streams: ws://localhost:3112
[OTel] Traces initialized: engine=ws://localhost:49134, service=agentmemory
[OTel] Metrics initialized: interval=30000ms
[OTel] Global fetch instrumentation enabled
[OTel] Logs initialized: delay=100ms, batch=1
[agentmemory] Consolidation pipeline: registered (CONSOLIDATION_ENABLED=false)
[agentmemory] Auto-compress: OFF (default, #138) — observations indexed via zero-LLM synthetic compression. Set AGENTMEMORY_AUTO_COMPRESS=true to opt-in to LLM-powered summaries (uses your API key).
[agentmemory] WARNING: AGENTMEMORY_INJECT_CONTEXT=true — the PreToolUse and SessionStart hooks will inject up to ~4000 chars of memory context into every tool turn. On Claude Pro this burns session tokens proportional to your tool-call frequency (see #143). Set AGENTMEMORY_INJECT_CONTEXT=false to disable.
[agentmemory] v0.6 advanced retrieval: sliding-window, query-expansion, temporal-graph, retention-scoring
[agentmemory] Orchestration layer: actions, frontier, leases, routines, signals, checkpoints, flow-compress, mesh, branch-aware, sentinels, sketches, crystallize, diagnostics, facets
(node:476815) [DEP0040] DeprecationWarning: The `punycode` module is deprecated. Please use a userland alternative instead.
(Use `node --trace-deprecation ...` to show where the warning was created)
[iii] Worker registered with ID: 683ce8e0-136a-43f7-8a28-78790cd6e9e5
[OTel] Connected to engine at ws://localhost:49134/otel
[agentmemory] Ready. Triple-stream (BM25+Vector+Graph) search active.
[agentmemory] Endpoints: 107 REST + 51 MCP tools + 6 MCP resources + 3 MCP prompts
[agentmemory] Auto-forget: enabled (every 60m)
[agentmemory] Lesson decay sweep: enabled (every 24h)
[agentmemory] Viewer: http://localhost:3113

[rohitg00]

You're right — my "rebuild the container" advice missed the actual root cause. Re-read the trace carefully now. Server is fine; the shim is what's degraded. Walking through what your output proves:

1. Server side: correct. curl http://localhost:3111/agentmemory/mcp/tools | jq '.tools | length' → 51. Server-side endpoint, AGENTMEMORY_TOOLS=all, healthy.

2. Shim side: silently in local mode. @agentmemory/mcp v0.9.6 has a two-mode flow (src/mcp/rest-proxy.ts):

   • probe GET http://localhost:3111/agentmemory/livez (500ms timeout)
   • probe OK → proxy mode → forward tools/list to server → 51 tools
   • probe fails → local mode → return IMPLEMENTED_TOOLS (7) from src/mcp/standalone.ts:16

   You see 7 tools in Roo Code, so the shim instance Roo Code spawned is in local mode. The probe failed for that process — even though it works from your host shell.

3. Likely cause: Roo Code spawns the shim inside the VS Codeium Flatpak sandbox. Even though Flatpak normally shares the host network namespace, two things commonly break the probe:

   • VS Codeium's Flatpak manifest restricts --share=network or routes localhost to a sandboxed loopback
   • The 500ms probe timeout is aggressive for first-call cold-start inside the sandbox

   The probe error is currently swallowed (catch {} in rest-proxy.ts:39), so there's no log line telling you what happened — that's a shim DX bug I'll fix.

Confirm the diagnosis — run this from inside the Flatpak (not your host shell):

flatpak run --command=sh com.vscodium.codium  # or whatever your VS Codeium flatpak id is
# inside the sandbox:
curl -v --max-time 1 http://localhost:3111/agentmemory/livez

If that fails or hangs, the probe is being blocked at the sandbox layer.

Workarounds while I land a fix:

Option A — point Roo Code at a tunnel that the Flatpak can reach (replace localhost with the host's LAN IP, e.g. 192.168.x.x):

{
  "command": "npx",
  "args": ["-y", "@agentmemory/mcp"],
  "env": {
    "AGENTMEMORY_URL": "http://192.168.x.x:3111"
  }
}

Option B — give the Flatpak network permission explicitly (run on host):

flatpak override --user --share=network com.vscodium.codium

Then fully quit VS Codeium and re-add the MCP server in Roo Code so the shim re-probes. (Roo Code, like most MCP clients, calls tools/list once per session and caches the result; restarting Codeium isn't enough — the MCP server entry needs to be removed/re-added.)

What I'm going to land in a follow-up PR:

1. rest-proxy.ts — surface probe failures to stderr with URL + reason (not silent), bump probe timeout from 500ms to 2000ms.
2. Add AGENTMEMORY_FORCE_PROXY=1 env so sandboxed clients can skip the probe entirely.
3. README troubleshooting note for Flatpak / Snap / sandboxed MCP clients.

Separate issues you flagged (filing these so they don't get lost):

• Data dir empty after docker compose down — volume binding bug. Will open a separate issue.
• which: no iii in ... plus reinstall every npx run — @agentmemory/agentmemory not finding the iii CLI on PATH. Separate.

Sorry for the misdirected first reply. Reopen this if the workarounds don't get Roo Code to 51 tools — I'd rather keep one thread on the shim/probe bug.