github Python analyzed 1c2b543

talos-kernel/Talos

github

Self-hosted AI agent for terminal and Telegram, with Claude/Codex workers, live activity and a deterministic permission kernel.

maintainer
talos-kernel
licence
MIT
first seen
2026-08-28
last seen
2026-09-15
releases · 30d
16
short id

talos-kernel/Talos is an MCP server distributed on github, maintained by talos-kernel, tracked here since August 2026. It has shipped 17 releases and exposes 13 tools. Tools include agent_consult, ask_operator, delegate, delegate_agy, delegate_codex, delegate_steer, and 7 more. Its composite risk grade is D — an inferred review prompt computed from observed signals, not a verdict.

what we found

Reading the source raised 4 review prompts — dangerous code, exfiltration combo, purpose mismatch and tool shadowing — each a pattern worth a human look rather than a finding of fault.

Its 13 tools appear to reach shell / exec, network and secrets, inferred from tool names, descriptions and input schemas rather than from observed behaviour.

Since the previous scan on 2026-09-15: 6 tools added.

full security breakdown →
tools 13
  • agent_consult bounded advice from a second, operator-configured agent — data, never permission
  • ask_operator the one way it can ask you something on purpose
  • delegate a sub-run that can only read
  • delegate_agy the same confined worker's second backend (the agy CLI): same trust form and workspace, behind its own opt-in gates on agent and worker, no…
  • delegate_codex bounded implementation and independent review with Codex: the same kernel and sandbox, isolated auth, verified stream completion and origin…
  • delegate_steer a course correction into a *running* /background task — a framed turn it reads at its next step, never a new right
  • entity_status a known name resolved to an operator-configured fixed read-only probe
  • git clone / fetch / pull / push with credentials — every op a human's yes, binding to exact op+repo+remote; local git work stays in the sandbox…
  • http_request any REST API — read methods through the SSRF-hardened door, state-changing methods always a human's yes, binding to exact method+URL
  • remote_exec a command on another machine over ssh — operator-allowlisted hosts, always a human's yes, standing rules bind to exact host+command
  • run_shell a command, sandboxed, or refused where no sandbox exists
  • session_search what was said in earlier turns
  • skill_write a new skill, written exactly once — and never without a human's yes
release cadence · 90d 17 releases
06-18 00:00 08-01 00:00 now
recent releases last 10
version date src
v0.19.17-alpha 2026-09-15 github
v0.19.16-alpha 2026-09-15 github
v0.19.15-alpha 2026-09-13 github
v0.19.14-alpha 2026-09-13 github
v0.19.13-alpha 2026-09-12 github
v0.19.12-alpha 2026-09-12 github
v0.19.11-alpha 2026-09-12 github
v0.19.9-alpha 2026-09-12 github
v0.19.8-alpha 2026-09-12 github
v0.19.7-alpha 2026-09-12 github

view all →