symm-mcp

Symm is a lightweight asynchronous MCP channel for handing work between independent agents and re-entering it with fresh attention.

The work persists. The viewpoint changes.

Any MCP client can dispatch a task to another agent, keep working, and later observe the result and record a resolution. Symm does not decide who implements and who reviews; that is expressed in the prompt.

Symm sits between two protocols: toward its caller it is an MCP server, and toward the agents it launches it is an Agent Client Protocol (ACP) client.

+-----------+      +----------------------+      +--------------+
| Dispatch  | ---> | Observe              | ---> | Resolve      |
|spawn_task |      |get_events / get_task |      |resolve_task  |
+-----------+      +----------------------+      +-------+------+
       ^                                                 |
       |        Repeat: next task / new viewpoint        |
       +-------------------------------------------------+

Status

0.2.0, the latest release (changelog), published on PyPI as symm-mcp. The package is classified as Alpha.

How a Task Flows

spawn_task returns as soon as the agent process is running. ACP session setup and the prompt continue after that; the caller can do other work and later read only events it has not seen yet.

Caller                     Symm server                Agent
  | spawn_task(...)              |                       |
  |----------------------------->| start process group   |
  |<-- task_id, status=running --|                       |
  |                              | initialize, new or    |
  |                              | load session, settings|
  |                              |---- prompt over ACP ->|
  |                              |<- updates/permission -|
  | wait_tasks([task_id])        |                       |
  |----------------------------->|                       |
  |                              |<-- turn ends ---------|
  |<-- completed, terminal ------|                       |
  | get_events(after_seq=0)      |                       |
  |----------------------------->|                       |
  |<-- page, cursor, has_more ---|                       |
  | get_events(after_seq=cursor) |  ... until has_more   |
  |----------------------------->|      is false         |
  | resolve_task(...)            |                       |
  |----------------------------->|                       |
  |<-- task with resolution -----|                       |

The next task may swap the agent and caller roles.

Two Independent Axes

Execution status says what happened to the process. Resolution says what a caller concluded about the result. They never change each other: succeeded does not mean accepted, and a resolution can be revised later.

Execution status (owned by Symm)

created --> starting --> running --success--> succeeded
  |           |             |-- failure --> failed
  |           |             +-- cancel/shutdown --> cancelled*
  |           +-- start failure -----------------> failed
  |           +----------------------------------> cancelled
  +----------------------------------------------> cancelled

* If process-group cleanup cannot be confirmed, the task ends as failed.

Resolution (recorded by callers, finished tasks only)

unresolved --resolve_task--> judged
judged --resolve_task--> judged
judged = accepted | rejected | needs_followup | superseded

For a clean completion, succeeded means the agent's prompt turn ended with stop reason end_turn. This does not show that the agent did useful work: check for agent_message or tool_call events, or verify the work, before accepting it.

Symmetric by Design

There is no built-in implementer or reviewer role. The same eight tools cover every direction work can travel:

Implement, then verify yourself
  You -- implement X --> Agent
  You <-- result ------ Agent

Implement yourself, ask for review
  You -- review workspace --> Agent
  You <-- findings ------- Agent

Chain or fan out
  Agent A -- implement --> Agent B -- find flaws --> Agent C
      |-- review 1 --> Reviewer 1
      +-- review 2 --> Reviewer 2

initiator_id and executor_id are descriptive labels only. They never grant ownership or authority.

Tools

Tool	What it does
describe_agent	Reads advertised model/effort offers as {current, values}; launch flags have a null current, and effort values describe the default model. Process startup and ACP setup share a 30-second deadline; ending the process group afterward is a separate, bounded supervisor cleanup attempt. Creates no task.
spawn_task	Dispatch a prompt to a registered agent. Returns as soon as the agent process is running, with the task's notify_path.
get_task	Current state: execution status, exit_code, error (the latest error message, usually why it failed; one recorded after the task ended does not replace it), session_id, resolution, timestamps.
get_events	One bounded page of events after after_seq; returns cursor, has_more, status.
wait_tasks	Wait up to 30 seconds for any given task to end; returns summaries and a reason.
list_tasks	This server's tasks, newest first, as summaries; recovers task ids a caller lost.
resolve_task	Record accepted, rejected, needs_followup, or superseded for a finished task. No side effects; may be revised.
cancel_task	Stop a running agent's process group; failed cleanup ends the task as failed. A finished task is unchanged.

An agent may keep the session it created for a description.

Registered agents. Symm is the Agent Client Protocol (ACP) client for six agents, and drives every agent over ACP. The Claude Code and Codex adapters launch the user-installed claude-agent-acp and codex-acp executables. Copilot, Cline, Cursor, and opencode are launched through their ACP commands.

The caller's model and effort choices reach each agent through the interface it accepts. Claude Code uses ACP session settings model and effort. Codex and Copilot use model and map effort to the session setting reasoning_effort. Symm checks session settings against the agent's ACP offer before sending the prompt. Cursor and opencode accept only model as a session setting. Cline takes model as a session setting and passes effort as the launch flag --thinking=<effort>.

Agent	Executable Symm launches	options
claude_code	claude-agent-acp	allow, model, effort
codex	codex-acp	allow, model, effort
copilot	copilot --acp	allow, model, effort
cline	cline --acp	allow, model, effort
cursor	cursor-agent acp	allow, model
opencode	opencode acp	allow, model

Permissions

Symm never widens an agent's permissions on its own: it approves only request kinds in allow, never chooses an "always" option, and pins the permission settings the caller cannot set. Actions an agent takes without sending Symm a permission request are outside what Symm observes or enforces. Broader enforcement, such as sandboxing, belongs to the caller or an external policy layer.

• Symm answers agent-sent permission requests. The allow option lists ACP tool kinds approved for this task: read, edit, delete, move, search, execute, think, fetch, and other. For an allowed kind, Symm chooses a one-time allow option; otherwise it chooses a one-time reject option. It never chooses an option that applies always. Mode-switch requests are never granted. If the one-time option it needs is not offered, Symm cancels the request. A decision Symm cannot record is cancelled, and the task fails. Each prompt-turn decision is recorded as a permission event. A request outside the prompt turn is cancelled without an event.
• Permission settings are pinned. Before the prompt, Symm sets Claude Code mode=default, Codex mode=read-only, Copilot its agent mode and then allow_all=off, and Cline auto_approve=false. Copilot's autopilot mode turns allow-all on and continues the turn on its own, and a resumed Copilot session keeps its saved mode. Cursor is pinned to mode=agent only when allow contains edit; otherwise its mode is plan. The opencode adapter is pinned to mode=build and configured to ask before edits, shell commands, and web fetches. Callers cannot set these pinned settings. If an ACP agent does not offer a setting or value Symm must pin, the task fails before the prompt and records an error event naming it. The opencode adapter's ask rules are supplied through OPENCODE_CONFIG_CONTENT: Symm preserves the user's other inline configuration, overrides those permission entries, and rejects a value that is not a JSON object. An empty value, or one of only JSON whitespace, counts as unset.
• Cursor has a mode-specific limitation. Symm pins Cursor to plan when allow lacks edit; in plan mode Cursor itself writes no files and runs no commands. In agent mode Cursor creates and edits files without asking, so Symm selects it only when allow includes edit. Cursor labels file deletions as edit, so allowing edit also allows deletion.
• Launch flags are checked by the agent. Cline's effort launch flag accepts none, low, medium, high, or xhigh; a value it rejects fails the task. Model and effort session settings are checked against the agent's offer before the prompt.

Events

Every created task records task_created; each status transition is a status_changed event. When the agent process starts, Symm records process_started; after supervision finishes, it records process_exited. A start failure records error without a process start or exit. Depending on what happens, a task can also record task_cancelled and task_resolved. The event types a task records today are: task_created, process_started, stderr, status_changed, process_exited, task_resolved, task_cancelled, error, session_started, agent_message, agent_thought, tool_call, tool_call_update, plan, permission, and agent_update. New event types may be added in any release. Consumers must ignore event types they do not know.

process_exited carries the agent's exit_code and the turn's stop_reason; with no requested stop or cleanup failure, only end_turn counts as succeeded, regardless of process exit code. A requested stop ends as cancelled even when the agent reports end_turn.

session_started records the agent session id, stores it in the task's session_id, and makes it available to spawn_task for a later resume. All six adapters support resume. A resumed task starts a new agent process and loads the existing session in the same workspace. It records only updates from the new prompt turn, not history replayed while loading. It uses only the model and effort choices supplied to that spawn_task call; Symm does not reapply choices from an earlier task. An agent that cannot load sessions fails before the prompt. If it rejects or does not recognize an id, the error event names the id and says it may not exist for that agent or workspace.

Each ACP request before the prompt (initialize, new or load session, and setting changes) has a 120-second timeout. If the agent does not answer, the task fails before the prompt and the error event names the unanswered request, such as load_session. An invalid or unavailable model/effort session setting fails before the prompt and records an error naming the setting; it also names the value when the agent offers that setting but not that value.

ACP message and thought chunks are coalesced and recorded at least once per second. tool_call and tool_call_update omit content blocks; raw output whose JSON encoding exceeds 4,000 characters is truncated to a preview. An agent's stdout carries the protocol and is not recorded; its stderr lines are recorded as stderr events.

Symm's own invalid-request and lookup or transition failures are MCP tool errors whose text has the form <code>: <message>. Codes are unknown_agent, invalid_request, task_not_found, and invalid_transition. The MCP SDK rejects arguments that do not fit a tool's input schema, including wrong types, missing required arguments, and unsupported values such as a resolution outside its allowed values, before Symm handles the call. These validation errors use the SDK's error text and have no Symm error code. Failed starts return a failed task with an error event. ACP session and supervision failures may also record error; an agent that exits before ending its turn fails the task.

Use

Requires uv, plus the CLI of each agent you use on PATH (claude-agent-acp, cline, codex-acp, copilot, cursor-agent, opencode). Symm does not install, update, or document the installation of agents: install and authenticate each one following its vendor's instructions (intent/agent-adapters.md). Symm runs the first match for each executable name in the absolute entries of the server's own PATH; which copy and version that is, is up to you.

Register Symm with an MCP client, for example Claude Code:

claude mcp add symm -- uvx symm-mcp@0.2.0

Or in a client's JSON configuration:

{
  "mcpServers": {
    "symm": {
      "command": "uvx",
      "args": ["symm-mcp@0.2.0"]
    }
  }
}

A typical loop, as the calling agent sees it:

1. spawn_task(agent="claude_code", prompt="Review the authentication changes in this workspace and list concrete flaws.", workspace="/path/to/repo") returns a task_id while the reviewer starts working.
2. Continue with other work.
3. wait_tasks([task_id]) until it returns reason terminal with the task in its completed list. Each call waits at most 30 seconds. Then read its events with get_events(task_id, after_seq=<cursor>), passing the returned cursor back while has_more is true; event_types narrows what is returned.
4. resolve_task(task_id, resolution="needs_followup", note="two findings to fix"), then dispatch the next task, possibly with the reviewer and implementer roles swapped.

Waiting Without Polling

spawn_task returns notify_path. When a task ends while the server is running, Symm creates an empty file at that path. If the server shuts down first, it removes the directory and no marker appears; read the outcome with get_task. A caller whose host wakes it when a background command ends, as Claude Code does, can run this in the background instead of calling wait_tasks again and again:

until [ -e "$P" ] || [ ! -d "$(dirname "$P")" ]; do sleep 1; done

P is the notify_path. The loop checks before it waits, because a task can end before spawn_task returns. The file only signals that the task ended. Read the outcome with get_task. If the directory is gone, the server has usually exited. Call get_task anyway: if the server still answers, its base directory was moved, and you can fall back to wait_tasks. A server killed with SIGKILL, or one whose file system stalls during shutdown, can leave the directory behind, so give the loop your own deadline if the wait must end.

Markers live in a directory of their own for each server process, inside a per-user base: $XDG_RUNTIME_DIR/symm-mcp, or symm-mcp-<uid> in the system temporary directory, where <uid> is the effective uid. Pass --notify-dir DIR to choose another base, or --no-notify to turn markers off, after the package in the registration above, for example symm-mcp --no-notify. If the base is a symbolic link, is not yours, or is writable by others, Symm logs a warning and starts with markers off. notify_path is null when markers are off, whether because of --no-notify or because Symm refused the base at startup (intent/completion-markers.md).

Agents run in the given workspace with the server's environment plus values their adapter sets; adapter values override inherited ones. The opencode adapter rewrites OPENCODE_CONFIG_CONTENT to include its permission configuration. Pass workspace explicitly: the default is the server's working directory, which your MCP client chose and may be your home directory. Symm does not copy, isolate, or clean workspaces.

Architecture

MCP client
   | stdio
   v
+------------------------ symm-mcp process -------------------------+
| MCP server (eight tools, errors, signals)                         |
|    |                                                              |
|    v                                                              |
| TaskService (orchestration, per-task locks)                       |
|    +--> Task store (JSON strings in memory)                       |
|    +--> Agent registry (adapter option allowlists)                |
|    +--> TaskSupervisor (process groups, argv without a shell)     |
|              |                                                    |
|              +-- launches ACP agents in shared workspace          |
|              +-- reads agent stderr, observes exits               |
|              +-- runs the ACP session client                      |
|                    (sessions, settings, permissions)              |
|                         |                     ^                   |
+-------------------------|---------------------|-------------------+
                          | ACP over stdio      | session updates,
                          v                     | permission requests
                 ACP agent processes (own process groups)

Logical state (tasks, events, resolutions) lives in the store and is always serializable. Runtime state (processes, pipes, process groups) lives only in the supervisor. The MCP layer stays a thin transport. These boundaries are part of the project contract in PROJECT.md.

Scope and Limitations

• stdio only, process-local task ledger. Each MCP client launches its own Symm server, and each server sees only the tasks it dispatched. stdio + in-memory = process-local task ledger.

  Agent A
     | stdio
     v
  Symm for A [ledger: task 1]
     | task 1
     v
  Agent B
     | stdio
     v
  Symm for B [ledger: task 2]
     | task 2
     v
  Agent C
  

  Agent A sees task 1 but not task 2, which B dispatched through its own server. A shared ledger needs an HTTP gateway with a shared store, which is deferred.

• In memory only, by design. Task state is lost when the server process exits, and Symm does not persist it. The agent keeps its own session; to continue later, even from a new client or server process, keep the task's session_id and pass it to spawn_task. spawn_task usually returns before an ACP session exists, so read session_id with get_task or from the session_started event. Resume with the same agent and workspace. The ledger and earlier task's events do not resume with it.

• No total output cap or retention policy. Stderr lines longer than 1 MiB are split into pieces no larger than 1 MiB. There is no cap on total output; output remains in the process-local ledger, so long, high-output tasks grow server memory.

• Tools only. No MCP resources or subscriptions yet. Wait for completion with wait_tasks, which returns within 30 seconds, or in the background on the task's completion marker. Read progress with get_events and its cursor.

• Bounded process cleanup. Before reporting completion, Symm tries to end every remaining member of the agent's process group. When canceling a running ACP task, it first asks the agent to cancel its prompt turn, then uses SIGTERM and, if needed, SIGKILL with bounded waits. If Symm cannot confirm the group is empty, the task ends as failed and the supervisor stops managing it; a process can remain alive. On client disconnect or handled SIGTERM, SIGINT, or SIGHUP, the server applies the same bounded cleanup before it exits. A task requested after shutdown begins fails rather than starting. A server killed with SIGKILL or one that crashes cannot run its shutdown handling, so its agents may survive.

• Your environment, your permissions. Symm does not install agent software or provide an operating-system sandbox for agent actions. Dependencies, secrets, sandboxing, and workspace cleanliness are the caller's responsibility. Symm pins adapter-controlled permission settings and answers the ACP permission requests an agent sends according to allow.

• POSIX only. Process groups and signals; Windows is not supported.

Rationale: PROJECT.md and the decisions under intent/. Security assumptions and how to report a vulnerability: SECURITY.md.

Run Locally

From a checkout:

uv run symm-mcp

The command speaks MCP over stdio, so it is meant to be launched by an MCP client rather than used interactively.

Development

This project records its decisions as plerosis intent under intent/; read AGENTS.md before changing anything. The Definition of Done:

uv sync
uv run python -m compileall -q src tests
uv run python -m pytest
uv run ruff check .
uv run ruff format --check .
./scripts/changelog-guard.sh
plerosis check --run

License

Licensed under either of Apache-2.0 or MIT, at your option.