ComputerPilot MCP intentionally exposes powerful local capabilities including filesystem mutation, process execution, browser automation, durable jobs, Git operations, and Windows desktop automation when supported.
Treat access to the MCP endpoint as equivalent to granting a development agent access to the OS account that runs the server.
Please do not open a public issue, discussion, or pull request for a suspected vulnerability that could expose credentials, allow unintended local execution, bypass a safety boundary, or compromise release/runtime verification.
Preferred reporting path:
- Use GitHub Private Vulnerability Reporting / Security Advisories for this repository if available.
- If private reporting is unavailable, contact the maintainer through the GitHub profile without publishing exploit details.
Include:
- affected project version/commit;
- operating system and architecture;
- Python version;
- affected start mode (
local-httpor tunnel); - minimal reproduction steps;
- expected vs actual behavior;
- sanitized logs or evidence.
Never include real API keys, tokens, cookies, private source, or other credentials.
The server is not a sandbox around the OS account running it.
Depending on the active tool profile and platform, an authorized MCP client may be able to:
- read and modify files accessible to the account;
- start or terminate processes;
- execute commands;
- interact with browsers;
- create durable jobs and workflows;
- inspect system metadata;
- use Windows native input/UI Automation.
When external extensions are enabled (MCP_EXTENSIONS_ENABLED=true), reviewed provider executables run with the permissions of the OS user running the server.
- Not an OS sandbox: Trust classes, environment filtering, and process group / Job Object ownership restrict communication channels and manage process trees, but they do not provide an OS-level sandbox.
- Environment isolation: Providers receive a minimal private environment. Full parent environments, wildcards, and arbitrary paths are not passed. Credential references are resolved in private memory and scrubbed before logs, audits, health snapshots, and artifacts.
- See Extension authoring contracts and Agent contract for full trust policy definitions.
Run it only under an account whose local permissions match the access you intend to grant.
The local Streamable HTTP endpoint and control panel bind to loopback by default.
Do not directly expose:
127.0.0.1:8765/mcp127.0.0.1:8766/
to an untrusted network.
The control panel adds host/origin checks and a session control token for guarded local mutations, but it is still intended for local operator use.
Tunnel credentials should be supplied through environment variables or:
.secrets/control_plane_api_key.txt
.secrets/ and local .env files are ignored by Git.
Do not put credentials in durable job idempotency keys, Git commit messages, issue reports, workflow descriptions, or other metadata that may be retained.
The repository and project release archives do not vendor tunnel-client or Cloudflared executables.
Tunnel mode uses scripts/tunnel_runtime.py to obtain the official openai/tunnel-client runtime. The updater validates the selected release asset, SHA-256 data, archive paths, symlink policy, and reported binary version before publishing it into .agent_state/tunnel-runtime/.
A fresh installation fails closed if no verified runtime can be obtained. An already-verified managed cache may be reused when a later update check is unavailable, according to the configured update policy.
See BINARY_PROVENANCE.md.
Generated state under .agent_state/ can include:
- durable job/workflow metadata and output;
- screenshots;
- recoverable edit backups;
- search snapshots;
- structured audit records;
- supervisor logs;
- recovery evidence;
- the managed tunnel runtime.
Project memory records live under memory/.
Do not publish these directories as release artifacts or attach them wholesale to issues. Sanitize logs and inspect screenshots/output before sharing.
Persistent language servers and installed package-test runners execute reviewed local code with the account's permissions; executable hashes, fixed protocol methods, environment filtering and process ownership are not an OS sandbox. Discovery never authorizes automatic installation. Keep provider configuration private and review the complete dependency supply chain.
Native cooperative resource claims fence managed edits, jobs, workflows, analyzers and file batches. They do not lock out independent editors or unmanaged external programs. After timeout/crash, uncertain ownership remains quarantined until evidence-based resolution; do not delete SQLite rows or backups to make admission succeed. Before the 0.8-to-0.9 schema upgrade, drain owners and take consistent state backups; an older runtime must use its matching pre-upgrade backup.
Arbitrary terminal commands (run_process, run_shell, run_cmd, run_powershell) take exclusive host coordination and a working-directory write claim. The optional access="read" field is accepted for wire compatibility and reported as a hint only; it never lets arbitrary shell code run concurrently, because a command cannot be proven read-only from its input. Native typed readers retain their read concurrency. This is a correctness and resource-safety boundary, not an authorization boundary: any client with MCP access can still execute commands permitted by the server account.
cancel_operation(operation_id) is an explicit, process-local cancellation contract for in-flight
terminal commands. The identifier is a bearer capability, not client authentication. A client
that can reach the same MCP server and learns an active id can request cancellation; the server does
not derive a separate client identity from loopback HTTP or stateless streamable HTTP. Server-generated
ids use 128 bits of randomness, client ids must be 8–128 characters, and ids are not persisted, so a
restart reports operation_not_found. Cancellation audit records contain only a process-local keyed
correlation, never the bearer id. Do not put ids in shared logs or expose the MCP endpoint to
untrusted clients. This is not an authorization boundary: any client with MCP access can already
invoke the full tool surface. Durable jobs use the separate cancel_job(job_id) contract and are not
cancelled by cancel_operation.
The runtime deliberately avoids blind replay of ambiguous mutations.
When an operation may have produced an external side effect but its final result was not durably recorded, it can remain uncertain until reconciled from evidence or explicitly acknowledged by an operator.
This reduces duplicate side effects but does not turn arbitrary external commands into transactions.
Release packaging is generated from tracked Git blobs and rejects forbidden runtime/secret/build-state paths. Published releases include SHA256SUMS.txt.
GitHub release publication is gated on the repository's CI, platform validation, and packaging workflows for the same release SHA.
These controls reduce accidental release drift; they are not a formal security certification or guarantee.