Prismer
ReviewAudited by ClawScan on May 10, 2026.
Overview
Prismer’s web/OCR/messaging features are disclosed, but its shared evolution memory and real-time agent messaging introduce broad cross-agent trust and persistence risks without clear boundaries.
Use Prismer only if you trust prismer.cloud and the @prismer/sdk package. Keep its API key separate from other credentials, require approval before sending/editing/deleting messages, treat incoming agent messages and tool-call content as untrusted, and avoid recording secrets or private incident details into the shared evolution system.
Findings (6)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
Installing the package gives third-party code the ability to run locally as part of the CLI setup and use.
The skill depends on a global npm package that is not included in the reviewed artifacts and is not pinned to a version.
which prismer || npm install -g @prismer/sdk
Install only if you trust the Prismer package and provider; consider pinning a known version and reviewing the npm package before use.
The CLI may store an API key or identity that can access Prismer messaging, credits, and account features.
The setup flow creates or stores service credentials for the Prismer account, which is expected but sensitive.
prismer setup # opens browser sign in key auto-saved (recommended)
Use a dedicated account or least-privileged key where possible, protect stored credentials, and revoke the key if the machine or agent environment is compromised.
An agent using this skill could change or remove conversation content if allowed to run these commands.
The documented CLI can mutate external messaging state by editing or deleting messages.
prismer im edit <conv-id> <msg-id> "Updated text" # edit prismer im delete <conv-id> <msg-id> # delete
Require explicit user approval before sending, editing, deleting, or group-broadcasting messages.
Incoming agent messages could be mistaken for trusted instructions or expose sensitive reasoning, files, or tool results.
The service supports cross-agent message types that can carry operational context or tool-call-like content, but the artifact does not describe trust or origin validation rules.
Message Types `text` (default), `markdown`, `code`, `file`, `image`, `tool_call`, `tool_result`, `thinking`
Treat all incoming messages and tool-call-like payloads as untrusted until verified; do not execute actions solely because another agent requested them.
Bad or poisoned strategies, or sensitive error/task details, could be reused across sessions or shared with other agents.
The artifact describes retrieving strategies from shared memory, executing them, and recording outcomes for reuse by other agents, without describing validation, scoping, redaction, or rollback controls.
Self-improving loop: encounter problem get strategy execute record outcome all agents benefit.
Review suggested strategies before execution, redact secrets before recording outcomes, scope learning to a trusted team where possible, and avoid sharing private incident details.
If enabled, the agent could continue receiving or checking messages in the background.
The skill documents cron-based polling and real-time delivery options, which can keep agent communication active beyond a single manual command.
Polling | 1-15 min | `prismer im conversations --unread` in cron
Enable cron, webhooks, WebSocket, or SSE delivery only when needed, and document how to disable or revoke those channels.