Cortex

An evidence-guided agent kernel for software-engineering agents.

Cortex is a small, local-first runtime that sits between an LLM and a set of specialist tools. It does not replace the model's planning or coding ability — it supplies the parts models are consistently bad at preserving across long tool-using tasks:

• a stable task state,

• explicit evidence and uncertainty,

• disciplined tool selection,

• bounded changes,

• verification tied to user-visible behavior,

• durable, structured memory,

• secret-safe execution.

The result is a model that is forced into a better reasoning loop:

orient → investigate → form hypotheses → declare a boundary → change → verify → preserve evidence

More tools without structure = more ways to get lost. Specialized tools + a kernel = accumulated engineering judgment.

Why

A capable coding model often fails not because it cannot write a function, but because it cannot maintain epistemic discipline through a multi-step task:

user report → broad search → read a few files → infer a likely cause →
edit several files → run one command → interpret partial success as proof →
lose the trail of why each decision was made

Cortex makes tool use stateful, bounded, inspectable, and evidence-driven — without becoming an autonomous black box. The model stays the planner and author; Cortex ensures its claims have an evidence trail and its actions occur inside a declared operating envelope.

Related MCP server: Axiom-hub

What it does

Cortex exposes six cognitive actions instead of dozens of overlapping raw tools:

These are structural invariants enforced by a phase machine, not by prompting. A model can't skip the disproof path by restating a hypothesis, or call a change "done" without proof.

Install

go install github.com/abdul-hamid-achik/cortex/cmd/cortex@latest
# or, from a clone:
task build   # → ./bin/cortex

Cortex is a single pure-Go binary (CGO_ENABLED=0). The specialist tools it composes (codemap, vecgrep, cairn, glyph, fcheap, vidtrace, tvault) are optional at runtime — every adapter degrades safely when its tool is absent, and Cortex never fabricates a missing tool's output.

Quick start (CLI)

# 1. Open a case and orient
cortex start "Fix post-login checkout redirect" --surface code --surface browser
#   → task_06FK…  [investigating]

# 2. Investigate (routes vecgrep → codemap, records evidence)
cortex investigate task_06FK… "where is the OAuth return URL handled"

# 3. Plan — a hypothesis WITH a disproof path is mandatory
cortex plan task_06FK… \
  --hypothesis "returnTo is dropped before callback :: run login-from-checkout browser flow" \
  --file src/auth/callback.ts --symbol HandleCallback \
  --uncertainty "unsure whether state signing also strips it"

# 4. …make your edits within the boundary, then verify
cortex verify task_06FK… \
  --claim "the OAuth callback preserves the return URL" \
  --browser-spec specs/cairntrace/checkout_return.yml

# 5. Complete — needs a verification receipt to succeed
cortex remember task_06FK… "returnTo was dropped from signed state; fixed and browser-verified" \
  --tag auth --tag oauth

# anytime: inspect the case
cortex status task_06FK… --detail full
cortex list
cortex doctor              # environment + specialist tool health
cortex studio               # interactive read-only case browser (Charm v2 TUI)

Every read command supports --json for machine consumption. Output is styled at a TTY and plain when piped.

MCP server

Cortex speaks the Model Context Protocol over stdio (newline-delimited JSON-RPC):

cortex serve

It exposes ten tools: cortex_start_task, cortex_investigate, cortex_plan, cortex_verify, cortex_remember, cortex_status, cortex_resolve, cortex_abort_task, cortex_read_evidence, cortex_read_artifact.

Register it with mcphub:

mcphub add cortex cortex serve
mcphub sync --write

In gateway mode the agent sees only mcphub, which proxies Cortex's tools as cortex__<tool> and keeps the raw specialist tools available as an expert escape hatch.

The case file

Each non-trivial task gets a durable, human-readable case file under .agent/cases/<taskId>/:

.agent/cases/task_06FK…/
  case.json          # goal, workspace identity, phase, change boundary, required verification
  evidence.jsonl     # append-only ledger of claims with provenance and confidence
  hypotheses.json    # falsifiable explanations + disproof paths
  plan.json          # the planning gate
  verification.json  # receipts: which claim, which verifier, passed/failed/not_run
  commands.jsonl     # non-sensitive audit trail of tool invocations
  summary.md         # the readable outcome

It is the kernel's working memory, not a transcript. Cortex writes .agent/.gitignore so its own state never registers as a workspace change.

The ecosystem it composes

Documentation

Full docs (built with VitePress) live in docs/. Run locally with task docs. The design specification is in SPEC.md; architecture and contributor guidance in AGENTS.md.

Status

v0.1 (MVP). The kernel, both surfaces (CLI + MCP), all eight adapters, the case-file store, redaction, scope-drift detection, and the verification policy are implemented and tested. See SPEC.md §21 for the milestone roadmap.

License

MIT © 2026 Abdul Hamid Achik