SPM Agent Connectors
Public connector package for SPM - Structured Project Memory.
SPM turns project knowledge into durable, shareable, governed and internally
smart project memory for AI agents. This repository contains only the public
connector surface: the authenticated remote MCP configuration, installable
bundles for Codex, Claude Code, Cursor and OpenClaw, browser authorization
helpers, lifecycle adapters, examples and security guidance.
The SPM application, backend, billing system, private console, memory engine and
infrastructure are not included in this repository.
Hosted MCP Endpoint
Human setup and inspection guide:
https://getspm.com/agents
https://getspm.com/v1/mcp
Official MCP Registry descriptor:
Registry name:
The endpoint is public for discovery, but authenticated for use. A client needs:
- an SPM account;
- an organization and at least one project;
- a trial or paid plan;
- an SPM token authorized for one project, a selected project set or the
projects the user may access in an organization.
What Agents Can Do
Connector profiles expose the agent-facing SPM surface:
- use durable project memory across chats, runs and tools;
- resolve the active project without confusing it with other authorized
projects;
- list authorized projects and compose cross-project context only when the
user explicitly requests it;
- work with smart memory packs that preserve requirements, current decisions,
completed work, source-backed context and temporal signals;
- query temporal project memory;
- request scoped context packs;
- verify context pack provenance and hashes;
- query context graphs and boundaries;
- run policy-aware preflight checks;
- report completed actions, tests, evidence and decisions;
- submit agent lifecycle turns to LLM-first memory triage with source
provenance, when the client supports lifecycle hooks.
- choose selective, complete redacted, summaries-only or metadata-only source
capture independently from the memory that triage promotes for future use;
- inspect and verify an append-ordered capture journal without exposing retained
conversation bodies to agent tools.
- propose a source-grounded project-memory bootstrap when no authorized project
matches, request only the specific authorized evidence still needed, then wait
for the user to create, link or skip it in SPM;
- continue an authorized project in another agent or device through a
short-lived one-time handoff that carries references rather than memory bodies
or credentials;
- compare body-free workspace manifests before a handoff: Git identity,
revision and local-state hash; non-Git file or document snapshots; remote
version references; or memory-only work;
- refresh workspace manifests automatically at session start, before user
turns and after completed work when the client supports lifecycle hooks,
making stale observations and out-of-band changes visible without uploading
local file bodies;
- distinguish metadata-only source observations from material evidence and
request an authorized source handoff when current content is required.
Typical Use Cases
- Coding agents receive the project memory they should actually use before
changing code: requirements, current decisions, tests, constraints and
completed work.
- Delivery teams turn scattered project knowledge into scoped packs for
engineers, clients, partners or support without exposing the whole project.
- Partners or external collaborators receive bounded memory with source,
expiry, revocation, provenance and audit evidence.
- Agent workflows preserve memory after the task by reporting decisions, tests
and evidence back to SPM.
What Agents Cannot Do
The hosted MCP connector does not expose:
- billing or checkout;
- invoice payment;
- customer portal creation;
- destructive tenant administration;
- raw secrets;
- global operator tools.
This boundary is intentional. Agents receive the memory tools they need without
being given commercial or destructive administrative powers.
Codex
Install the Codex plugin from this repository marketplace:
.agents/plugins/marketplace.json
Or configure Codex manually:
[mcp_servers.spm]
url = "https://getspm.com/v1/mcp"
bearer_token_env_var = "SPM_CODEX_MCP_TOKEN"
startup_timeout_sec = 30
tool_timeout_sec = 120
Authorize Codex for all projects the signed-in organization user may access:
python3 plugins/spm-codex/scripts/auth_spm_codex.py \
--project-id <authorization-project-id> \
--access-mode organization \
--write-env ~/.spm/codex.env
source ~/.spm/codex.env
Verify:
python3 plugins/spm-codex/scripts/doctor_spm_codex.py
Run the functional remote MCP smoke:
python3 plugins/spm-codex/scripts/smoke_spm_remote_mcp.py \
--project-id <authorized-project-id>
The smoke uses the same SPM_CODEX_MCP_TOKEN as Codex and requires an
explicit authorized project id (or SPM_SMOKE_PROJECT_ID). It initializes the
hosted MCP endpoint, verifies the project-resolution, multi-project,
agent-session and memory tool surface, creates a small project-scoped smoke
memory event, reads temporal state, creates and verifies a context pack, queries
the context graph, runs agent preflight and reports post-action evidence. It
never prints the token and the hosted connector strips raw event bodies from
returned context.
For a non-mutating token check, use:
python3 plugins/spm-codex/scripts/smoke_spm_remote_mcp.py \
--project-id <authorized-project-id> --read-only
Claude Code, Cursor And OpenClaw
The repository includes native client bundles:
plugins/spm-claudeplugins/spm-cursorplugins/spm-openclaw
Clients that only support MCP can use the same remote endpoint and bearer
token. The generic authorization helper is
scripts/agent-connectors/authorize_spm_agent.py. See also:
examples/codex/config.tomlexamples/claude-desktop/claude_desktop_config.jsonexamples/cursor/mcp.jsonexamples/windsurf/mcp_config.json
Security
Do not commit tokens. Prefer environment variables or each agent client's
secret storage. Tokens must be explicitly scoped and revocable from SPM.
Organization visibility does not permit silent project mixing: each task keeps
one active project, while cross-project packs require an explicit user request.
Session continuation does not clone, pull, reset or overwrite a workspace.
Receiving agents inspect their actual material state, compare it with SPM's
manifest and ask before obtaining or reconciling resources.
See SECURITY.md and docs/security-boundary.md.
Links